UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Computational model of a behaviour in c. elegans and a resulting framework for modularizing dynamical… Roehrig, Chris J. 1998

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata


831-ubc_1998-346188.pdf [ 7.71MB ]
JSON: 831-1.0099377.json
JSON-LD: 831-1.0099377-ld.json
RDF/XML (Pretty): 831-1.0099377-rdf.xml
RDF/JSON: 831-1.0099377-rdf.json
Turtle: 831-1.0099377-turtle.txt
N-Triples: 831-1.0099377-rdf-ntriples.txt
Original Record: 831-1.0099377-source.json
Full Text

Full Text

A Computational Model of a Behaviour in C. elegans and a Resulting Framework for Modularizing Dynamical Neuronal Structures by Chris J. Roehrig B.Math (1988), M.Math (1991), University of Waterloo A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF Doctor of Philosophy in THE FACULTY OF GRADUATE STUDIES (Program in Neuroscience) We accept this thesis as conforming to the required standard The University of British Columbia September 1998 © Chris J. Roehrig, 1998 In presenting this thesis in partial fulfilment of the requirements for an advanced degree at the University of British Columbia, I agree that the Library shall make it freely available for reference and study. 1 further agree that permission for extensive copying of this thesis for scholarly purposes may be granted by the head of my department or by his or her representatives. It is understood that copying or publication of this thesis for financial gain shall not be allowed without my written permission. Department of ("jW^ dfel ^ W A ^ / k] ^MAQ^IS^A^ The University of British Columbia Vancouver, Canada Date od IH h i DE-6 (2/88) Abstract The work presented in this dissertation grew out of a study of a physiologically based computational model of the tap withdrawal response in the nematode Cae-norhabditis elegans. A computational model using all available anatomical and physiological data was unable to explain a dynamic property of the circuit: the ability of the behaviour to continue after the termination of the stimulus. To account for this behavioural observation, a novel approach was taken: a neuronal circuit was engineered from a set of modules each consisting of several physiologically realistic model cells. The mathematical dynamics of the resulting neuronal circuit produced an output that was similar to the behaviour observed in the intact worm and shows that neuronal network dynamics could account for the behaviour. In the course of this study, it became clear that little is known about the mod-ular properties of neuronal dynamics. This dissertation presents an approach for combining non-linear neuronal circuits into larger systems using dynamical mod-ules (dymods), and a set of tools for studying dymods, and discusses a research strategy for studying the modular properties of neuronal dynamics. ii Contents Abstract ii Contents iii List of Tables viii List of Figures ix List of Symbols xii Acknowledgements xv Dedication xvii 1 Introduction 1 1.1 Thesis Statement . . . . 4 1.2 Document Overview 4 1.2.1 A Physiological Model of C. elegans Cells 5 1.2.2 Dymods: A Framework for Modularizing Dynamical Neu-ronal Structures 6 iii 1.2.3 The DSS Network Protocol for Dymod Implementation . . 7 1.2.4 A Desktop Robot System for Experimental Neuroethology using Dymods 8 2 A Physiological Model of C. elegans Cells 9 2.1 Introduction 9 2.2 The Model 9 2.3 Neurons 13 2.4 Gap Junctions 18 2.5 Synapses 19 2.6 Synaptic Activation 21 2.7 Steady-state Potential • 22 2.8 Synaptic Parameters 24 2.8.1 Modelling Ascaris Monosynaptic Response 25 2.9 The Gearbox 31 2.10 Results and Conclusions 33 3 Dymods: An Approach to Modularizing Dynamical Neuronal Struc-tures 36 3.1 Methods and Results 37 3.1.1 The Cell Model 38 3.1.2 The Switch Dymod 41 3.1.3 The Oscillator Dymod 45 iv 3.1.4 The Charger Dymod 49 3.1.5 Dymod Assembly: The Reversal Maintenance Circuit . . . 52 3.2 Discussion 53 3.2.1 The Locus of C. elegans Reversal Dynamics 55 3.2.2 Oscillations 58 3.2.3 The Forward-Engineering Approach 61 3.2.4 The Human Engineered Approach 63 3.2.5 Conclusions and Future Directions 66 4 The DSS Network Protocol for Dymod Implementation 69 4.1 Introduction 69 4.2 Design Goals 71 4.3 Review of Existing Frameworks 72 4.4 Theory of Operation 74 4.4.1 Signal and Time Representation 75 4.4.2 Signal Reconstruction . 77 4.4.3 Real-time Scheduling 80 4.4.4 Time Synchronization 83 4.4.5 Connection Management 85 4.4.6 Network Transmission 85 4.5 Conclusions 86 5 A Desktop Robot System for Experimental Neuroethology using Dy-mods 88 v 5.1 Introduction 88 5.2 Design Goals 89 5.3 Chassis Design 90 5.3.1 Locomotion 90 5.3.2 Proprioception and Sensors 92 5.3.3 Construction 94 5.4 Controller 94 5.4.1 Overview 95 5.4.2 BINMON: The Miniboard Control Program 96 5.5 Future Work 101 5.6 Conclusions 102 6 Conclusions 104 6.1 Future Directions 105 6.2 Novel Contributions 107 Bibliography 109 Appendix A The DSS Protocol Specification and API 118 A . l The DSS Protocol Specification 118 A. 1.1 DSS Port Addresses 119 A. 1.2 DSS Messages 120 A. 2 The DSS Application Programmer Interface (API) 127 Appendix B Robot Assembly Instructions and m b l i b API 132 B. l LEGO Tank Robot Construction 132 vi Miniboard Host C Library List of Tables 2.1 Average simulation parameters 13 2.2 Summary of neuron characteristics 15 3.1 Default model parameters 45 A . l DSS Message Header Flags 121 viii List of Figures 2.1 The complete connectivity of the tap withdrawal circuit 12 2.2 Example of neuron process length and diameters 16 2.3 Fit for the VI-VM synapse 29 2.4 Fit for the DE1-DM synapse 30 2.5 Comparison of model data to behavioural data 34 3.1 A coupled pair of cells 39 3.2 Nullclines for a coupled pair of cells 41 3.3 Fixed points for an inhibitory switch 42 3.4 Inhibitory switch turning on 43 3.5 Simulated recording traces for an inhibitory switch 44 3.6 Oscillator circuit 46 3.7 Delay due to increased steepness of activation curve 47 3.8 Simulated recording traces for an oscillator 48 3.9 Phase portrait of an oscillating switch 49 3.10 Phase portrait of charging behaviour 50 3.11 Simulated recording traces of charging circuit 51 ix 3.12 Complete reversal maintenance circuit 52 3.13 Simulated recording traces for the complete reversal circuit 54 3.14 The dymod view of the reversal maintenance circuit 55 4.1 A neuronal dymod control system for C. elegans reversal maintenance 70 4.2 Dymod interconnections in a coupled two-dimensional system . . . 74 4.3 Reconstruction Error vs Filter Width 79 4.4 DSS Reconstruction Errors vs Sampling Interval 84 5.1 Preliminary Tail-dragger Robot Design 91 5.2 LEGO Tank Robot Design 92 6.1 The Next Step in the Research Programme 106 A . l DSS Architecture 119 A.2 DSS Address 119 A.3 DSS Message Header 121 A.4 DSS Isochronous Message 122 A.5 DSS Asynchronous Message 122 A.6 DSS Connection Request Message 123 A.7 DSS Disconnection Request Message 123 A. 8 DSS Add Target Message 124 A.9 DSS Delete Target Message 124 A. 10 DSS Name Register Message 125 A. 11 DSS Name Query Message 125 A. 12 DSS Name Response Message 126 x A. 13 DSS Epoch Message 126 A. 14 Epoch Renegotiation 128 B. l Robot Tank Chassis LEGO Assembly, Part 1 133 B.2 Robot Tank Chassis LEGO Assembly, Part 2 134 B.3 Robot Tank Chassis LEGO Assembly, Part 3 135 B.4 Robot Tank Chassis LEGO Assembly, Part 4 136 B.5 Robot Tank Chassis LEGO Assembly, Part 5 137 xi List of Symbols Cm Membrane capacitance 17 d Process diameter 16 Esm Equilibrium (reversal) potential for synaptic channels 19 EsYNij Synaptic reversal potential for current Uj 23 EACTJ Presynaptic activation potential for cell j 40 ELEAK Membrane leakage equilibrium potential 17 ERANGE Presynaptic activation range (mV) 22 g maximal post-synaptic membrane conductance 22 cjij Gap junction conductance between cell i and cell j 18 9LEAK Membrane leakage conductance {giEAK =1/Rm) • 40 9OO(VPRE) Steady-state post-synaptic conductance 21 g(t) Synaptic conductance 19 gij Synaptic conductance to cell i from cell j 23 I m Externally injected current 17 IINJi Externally injected current into cell i 23 xii Iij Gap junction current into cell i from cell j 18 Iij Synaptic current flowing into cell i from cell j 23 ISYN Synaptic current 17 JijiVj) Dimensionless synaptic input to cell i from cell j 40 K Scaling constant for synaptic activation range 22 A Length constant 17 hij Number of gap junctions between cell i and cell j 23 n-ij Number of synapses to cell i from cell j 23 iV Number of cells in the circuit 23 Ri Intracellular resistivity (fi cm) 14 RM Specific membrane resistance (ktt cm2) 14 RM Membrane leakage resistance 17 RPOST Input resistance of postsynaptic cell 27 RPRE Input resistance of presynaptic cell 27 ^ Membrane time constant for cell i 39 Synaptic time constant for synapses to cell i from cell j 23 IINJ External input current in volts/sec 40 Vsst Steady-state in-circuit membrane potential for cell i 22 Vi Membrane potential for cell i 22 Vposr Membrane potential of postsynaptic cell 19 VpRE Membrane potential of presynaptic cell 21 xiii W{j Dimensionless weight for synapses into cell i from cell j 40 xiv Acknowledgements I'd like to thank my PhD supervisory and examination committees for their insightful remarks and guidance: Leah Keshet, Steve Kehl, Mark Greenstreet, Nick Swindale, and especially Jim Little and Cathy Rankin for footing the bill. In addition, my previous advisors Robert Miura, Bob Woodham, Dinesh Pai and the Computer Science Department's Laboratory for Computational Intelligence were instrumental in the inception of this work. I'd like to acknowledge Jennifer Enns-Ruttan, Bard Ermentrout and Randy Beer for helpful discussions in the construction of the dymod circuit of Chapter 3. Shawn Lockery provided useful advice and experimental evidence to support the physiological model, and Tom Ferree and Tom Morse gave feedback on the model and the robotic implementation tools. I owe thanks to Steve Wicks for starting on the path to modelling the C. elegans tap withdrawal circuit and working with me through to its publication, and Bruce Hutcheon for his participation and insight. Life would have been a lot tougher if not for my esteemed fellow House.ORGians: Tom Glenne, Yggy King, Brenda Mary Haggard, Hanan Elmasu, Vicki Sawyer, Sabine Tamm, Glenn Wells, and Jackie Taylor. A special thanks to Hugh Thompson for those late-night eye-scorchers that I hope will be remain the xv extremes of my terror, and Fiona Miller whose friendship made the first years here bearable. To Mike Horsch whose philosophical discussions have inspired me prob-ably more than anything else. To my family, Mom, Dad, Andy and Lisa for being there for me, and to my soulmate Christina Pechloff: your support made finishing this thing easier than I could have imagined. C H R I S J . R O E H R I G The University of British Columbia September 1998 xvi This thesis is dedicated to my supervisor, Cathy Rankin. Without her unfailing support through many trials, this thesis would never have been completed. xv i i Chapter 1 Introduction The brain has a rich dynamical structure. We already have a good idea of the complexity of its physical structure, its molecules, channels, cells, nerves, and nuclei, but we have only begun to understand the dynamical structure that emerges from the interactions among these components. In this thesis, the distinction is made between dynamical structure and dynamic structure, i.e. structure that is merely changing over time. An example of dynamic structure is an ion channel: a cell membrane protein that can change its molecular configuration to allow cur-rent to flow through. Rather, a dynamical structure may not exist as a physical entity but rather it only emerges from patterns of changes. For example, in pattern generators such as the lobster stomatogastric ganglion (Harris-Warrick, Nagy and Nusbaum, 1992), current flowing through ion channels can ultimately affect other connected cells and result in periodic patterns of activity — oscillations — among the cluster of cells. These oscillators are not physical structures: they exist only in the patterns of changes in the underlying physical structures. 1 Pattern generators like these are the simplest examples of dynamical struc-tures, but there are likely to be many levels of more complex dynamical structures. Patterns can emerge in the way that dynamical structures themselves change — pat-terns of change in the patterns of change, and so on. These dynamical structures do not necessarily correspond to any physical structures and may not be readily apparent by examining physical structure, as this thesis shows. There is a grow-ing appreciation among neuroscientists (Chiel and Beer, 1997) and cognitive and behavioural scientists (Port and van Gelder, 1995; Kelso, 1995) that understanding the dynamical structures that emerge from the brain's physical structure is the key to understanding how the brain generates behaviour. The dynamical study of neuronal circuits is still in its infancy, and has con-centrated on small systems. An interesting model system for studying small cir-cuits is the soil nematode C. elegans. Its nervous system contains only 302 neurons all of which have been completely described in terms of their location and synap-tic connectivity (White, Southgate, Thomson and Brenner, 1986; Hall and Rus-sell, 1991; Achacoso and Yamamoto, 1992). Moreover, the system is amenable to laser microsurgery so that it is possible to destroy individual neurons in the living animal with little or no damage to the remaining nervous system (Chalfie, Sulston, White, Southgate, Thomson and Brenner, 1985; Wicks and Rankin, 1995). In spite of this simplicity, it has a rich behavioural repertoire and has been chosen as a model system for studying learning and memory (Rankin, Beck and Chiba, 1990) and genetics (Wood, 1988). C. elegans has been studied in sufficient detail that it is feasible to construct a cellular account of a non-trivial behaviour. 2 One such behaviour that has been studied in detail is the tap withdrawal re-flex in which the nematode swims backwards in response to a vibratory tap stimulus to the agar jelly on which it locomotes (Rankin et al., 1990). The circuitry underly-ing this tap withdrawal response has been well-studied (Chalfie et al., 1985; Wicks and Rankin, 1995). It consists of a sensory and interneuronal subcircuit that pro-cesses potentially conflicting mechanosensory stimuli to arrive at a graded decision output (Rankin, 1991). Locomotion is mediated via two command interneuron pairs called AVA and AVB which drive two independent motor systems for reverse and forward sinuous locomotion respectively (Chalfie et al., 1985; White et al., 1986). We have previously constructed (Wicks, Roehrig and Rankin, 1996) a de-tailed computer model using all available anatomical and physiological data to predict the functional polarities of the synapses in the C. elegans tap withdrawal circuit. These predictions were made by optimizing the behaviour of a modelled circuit when various cells are removed to the tap withdrawal behaviour of real an-imals when the corresponding cells are removed using laser ablation. Our model accounts for how the tap response resolves the conflicting stimuli to head and tail to arrive at a decisive response, and when neurons are removed from the model, it responds in a way that is commensurate with the response of the animal when the corresponding cell is ablated. Although this model was useful for predicting the po-larities of the synapses in the circuit, it was unable to account for the continuation of the reversal behaviour when the stimulus ended. 3 1.1 Thesis Statement The hypothesis of this thesis is that network dynamics could account for re-versal maintenance. While the anatomical map of the nervous system is known, the connectivity alone is not sufficient to determine the dynamics of the neuronal circuitry. In-circuit recordings are not yet possible in the animal, and so the cellular patterns of activity during behaviour are still largely unknown. Therefore, a novel approach was taken: a neuronal circuit was engineered from a set of distinct mod-ules each consisting of several physiologically realistic model cells. The circuit was designed so that its mathematical dynamics produces an output that is similar to the behaviour observed in the intact worm. The novelty of this approach is the way that a human engineer was able to design the circuit dynamics in parts to simplify the process: the dynamics of each module were designed separately using model cells, and the modules were then assembled to form the complete circuit with the desired dynamical behaviour. The term "dymod" was coined to refer to these dynamical modules. 1.2 Document Overview This dissertation is organized into the following chapters. Chapter 2 de-scribes the physiological derivation of the mathematical and computational model used in our polarity predictions (Wicks, Roehrig and Rankin, 1996) and in the con-struction of the neuronal circuit to account for the reversal maintenance behaviour (Chapter 3). Chapter 3 describes the dynamical design of the reversal maintenance 4 circuit model and introduces the concept of dymods. Chapter 3 also discusses a research strategy for studying the interactions between dynamical neuronal struc-tures using dymods. Chapter 4 presents the Digital Signal Sockets (DSS) network protocol — a preliminary implementation framework for dymods that emphasizes their modular aspect by requiring dymods to have a clearly defined interface, and uses computer networks for interconnection. Finally, Chapter 5 presents another key component in the dymod research strategy: a desktop robot system for per-forming neuroethological experiments to investigate how dymods can be used to model complete behavioural systems. 1.2.1 A Physiological Model of C. elegans Cells Chapter 2 presents the derivation of the physiologically-based computational model used in Chapter 3 and previously used (Wicks, Roehrig and Rankin, 1996) to predict the functional polarities of the synapses in the C. elegans tap withdrawal circuit. The model uses a simple single-compartment model for the nearly isopo-tential cells and presents a novel graded and tonic synaptic model. The synap-tic model emphasizes the distinction between an isolated cell's leakage (resting) membrane potential and its in-circuit steady-state potential which differs because of tonic synaptic currents flowing in the circuit. The model's parameters were de-rived from all available anatomical and physiological data from C. elegans and a related nematode Ascaris. While the model successfully predicted the magnitude of the behavioural response in the animal, it did not account for the time course of the animal's reversal behaviour. To account for this discrepancy, a new model 5 that uses network dynamics to maintain the reversal was proposed, and this is the subject of the next chapter. 1.2.2 Dymods: A Framework for Modularizing Dynamical Neu-ronal Structures Chapter 3 presents the design of a circuit to help explain how the nematode C. elegans makes transitions between forward and reverse locomotory modes and maintains its reversal for a duration much longer than the time constants of the cells controlling the behaviour. This circuit shows that the maintenance of a reversal could be the result of network dynamics and a bifurcation between two quasi-stable states governing forward and reverse locomotion. To engineer the circuit, a novel approach was taken: a human engineer used intuition about the dynamical processes involved to hand-engineer a modular set of component dynamical systems ("dymods" — short for dynamical modules) that were assembled to form the final circuit. In spite of the non-linear interactions between dymods, the assembly was remarkably straight-forward suggesting that a human-engineered approach might not be so difficult as it seems. The reversal dynamics circuit designed in this chapter is a single continu-ous dynamical system consisting of simple tonic cells. Yet it governs two discrete behaviours and the crisp transitions between them. These behavioural transitions occur as the result of bifurcations in the underlying dynamical system and raises the question of whether bifurcations can account for other behavioural transitions. A research strategy for investigating the generality of this phenomenon is discussed. 6 1.2.3 The DSS Network Protocol for Dymod Implementation Chapter 4 presents the Digital Signal Sockets (DSS) network protocol as an experimental framework for interconnecting dynamical modules (dymods). Dy-mods are continuous dynamical systems with a denned functional interface of in-put and output signals designed for use in constructing complex neuronal models of behaviours for neuroethological experiments, but may have applications to other control systems. Dymods are implemented as numerical solutions to ordinary dif-ferential equations (ODEs) and DSS provides a standard mechanism to intercon-nect independent real-time dymod implementations together with each other and live motor and sensor signals. The DSS protocol solidifies the modular aspect of a dymod by providing an explicit definition of a dymod's interface inputs and outputs. DSS is targeted at two network architectures: TCP/IP for inexpensive exper-iments with low-bandwidth systems, and IEEE 1394 "Firewire" for high-bandwidth applications with guaranteed performance. This chapter discusses the following is-sues for a dymod implementation: signal and time representation, signal reconstruc-tion, connection management, time synchronization, real-time scheduling, and net-work transmission. The unoptimized experimental DSS implementation described here is suitable for low-bandwidth signals of less than 50 Hz and suggests the feasi-bility of using digital networks to interconnect real-time simulations of dynamical systems. 7 1.2.4 A Desktop Robot System for Experimental Neuroethology using Dymods Chapter 5 adds to the dymod tools by presenting a robot system for neu-roethology experiments. It continues the theme of the previous chapters with the goal of empowering the general neuroscience researcher, in this case to perform robot neuroethology experiments without requiring expertise in robotic engineer-ing or real-time numerical computation. The inexpensive LEGO robot uses a wheeled-design for simplicity and re-liability. It uses a tank-like chassis with treads, which gives it the ethologically relevant ability to orient in place, unlike other car-like designs which require three-point turns. The chassis is compact and houses the battery compartment, motors and computer control system below the tank's "deck" to provide maximum flexi-bility for adding sensory and actuator apparatus. The control system is a tethered design: an MIT Miniboard monitors and controls the robots sensors and motors and transmits them along a cable to a host UNIX computer which performs the actual neuronal computation. This chapter also presents a Miniboard program (BINMON) that performs the control functions, a UNIX library to communicate with the robot via serial cable, and a library add-on for the popular GENESIS neuronal simulator to allow it to communicate with the robot. 8 Chapter 2 A Physiological Model of C. elegans Cells 2.1 Introduction This chapter presents the derivation of the physiologically-based computa-tional model used in Chapter 3 and previously used (Wicks, Roehrig and Rankin, 1996) to predict the functional polarities of the synapses in the C. elegans tap with-drawal circuit. 2.2 The Model The nematode C. elegans reverses its locomotion and swims backwards in response to a tap stimulus (Rankin et al., 1990), and maintains this reversal for several seconds before resuming forward locomotion. The circuitry underlying this 9 tap withdrawal response has been well-studied (Chalfie et al., 1985; Wicks and Rankin, 1995). It consists of a sensory and interneuronal subcircuit that processes potentially conflicting mechanosensory stimuli to arrive at a graded decision output (Rankin, 1991). Locomotion is mediated via two command interneuron pairs called AVA and AVB which drive two independent motor systems for reverse and forward sinuous locomotion respectively (Chalfie et al., 1985; White et al., 1986). While the anatomical map of the nervous system is known, the connectivity alone is not sufficient to determine the dynamics of the neuronal circuitry. In-circuit recordings are not yet possible in the animal, and so the cellular patterns of activity during behaviour are still largely unknown. The model used was a physiologically motivated one. However, in the ab-sence of detailed physiological data from C. elegans, it was necessary to make a number of extrapolations from the related nematode Ascaris lumbricoides. These assumptions are presented in physiological rather than mathematical form to ensure that they are realistic. Furthermore, preliminary investigations suggested that po-larity predictions based on the modelled circuit were more strongly determined by circuit connectivity than the exact values of parameters used. Thus, approximate ranges for these parameters rather than precise values were derived. The effects of varying some of the more uncertain parameters were then assessed by rerunning the same experiments with the values of these parameters varied over three orders of magnitude. The circuitry was constructed by extracting connectivity data from AY's Neuroanatomy of C. elegans for Computation (Achacoso and Yamamoto, 1992). 10 This data indicated not only the presence or absence of a set of synaptic contacts between a pair of neurons, referred to here as a synaptic class, but also incorpo-rated the actual number of documented electrical and chemical connections within that synaptic class. Each synaptic contact within a class of synapses was assigned the same reversal potential and conductance as all other synapses within that class. This enabled the simple construction of complex circuits in which all documented synapses (including all bilateral asymmetries) were included in the model. It was assumed that the functional efficacy of a synaptic class was correlated with the number of contacts observed within that synaptic class. Thus, circuits constructed in this way possessed connections with weights determined by anatomical criteria. These weights were not varied further in this model; only the reversal potential, which determined the sign of the connection, was allowed to vary. The complete connectivity of the modelled tap withdrawal circuit is shown in Figure 2.1. The model was based on all available physiological and anatomical data from C. elegans and the related nematode Ascaris. The physiological parameters used in the derivation of the data presented in this report are shown in Table 2.1 and Table 2.2. The model was implemented in Objective-C on Intel-486, HP series 9000, and NeXT computers running NEXTSTEP software, and was integrated us-ing fourth-order Runge-Kutta (Press, Flannery, Teukolsky and Vetterling, 1988) to an accuracy of 0.5%. 11 Figure 2.1: The complete connectivity of the tap withdrawal circuit (extracted from Achacoso and Yamamoto, 1992). The circuit consists of seven sensory neurons (shaded circles), nine interneurons (unshaded circles), and two motorneuron pools (not shown), which produce forward and backward locomotion (triangles). Chemi-cal connections are indicated by arrows, with the number of synaptic contacts being proportional to the width of the arrow. Gap junctions are indicated by dashed lines. Every connection represented in this figure was also represented in the model. This representation is useful for identifying connection asymmetries which might un-derlie the origins of oscillations that control locomotion and are hidden in simpler views of the circuitry. 12 Neuron Parameters Value Units Membrane resistance see Table 2.2 Ohms Membrane capacitance see Table 2.2 Farads Membrane leakage potential -0.035 Volts Synaptic Parameters Value Units EPSP reversal potential 0.00 Volts IPSP reversal potential -0.048 Volts Synaptic conductance 6.00E-10 Siemens ERANGE 0.035 Volts Gap junction conductance 5.00E-09 Siemens Tap Parameters Value Units Pulse rest 0 Amps Phasic pulse 1.00E-11 Amps Start time 0.01 Sec Duration 0.3 Sec Tonic pulse 2.50E-10 Amps Table 2.1: List of physiological parameters used in the four experiments run in Wicks, Roehrig and Rankin (1996) are summarized. For a more detailed discussion on the origin of these values, see the text. 2.3 Neurons The neurons of C. elegans have simple morphologies which are preserved across individuals. Many neurons consist of a single unbranched process, and few have more than two branches (Wood, 1988). Electrophysiological analysis of C. el-egans cells is still in its infancy (however, see Raizen and Avery, 1994; Avery, Raizen and Lockery, 1995; Goodman, Hall, Avery and Lockery, 1998) and little is known about the membrane characteristics of its neurons. However, electro-physiology has been done on Ascaris lumbricoides, a larger nematode related to C. elegans (Davis and Stretton, 1989a; 1989b). Its dorsal and ventral nerve chords have been reconstructed and show considerable similarity to those of C. elegans, and anatomical homologues of C. elegans motorneurons have been found in Ascaris 13 (Wood, 1988; Stretton, Donmoyer, Davis, Meade, Cowden and Sithigorngul, 1992). For our model, we used electrophysiological data from Ascaris to determine our model parameters. Evidence from Ascaris suggests that signal propagation in C. elegans neu-rons is likely accomplished electrotonically, without classical all-or-none action po-tentials. Intracellular recordings of Ascaris motoneurons and interneurons show no evidence of action potentials, nor has it been possible to evoke them (Davis and Stretton, 1989a). Specific membrane resistance in Ascaris is unusually high and is within the range that would permit signal propagation without action potentials. Niebur and Erdos (1993) have used Ascaris data to do detailed computational stud-ies of the electrotonic characteristics of C. elegans neurons and have shown that the function of C. elegans locomotion circuitry can be accounted for by purely electro-tonic signals. Davis and Stretton (1989a) have measured specific membrane resistances R m and intracellular resistivity R{ in Ascaris. In four motorneurons, R m varied from 61 - 251 kfi cm2, and R, from 79 - 314 fi cm. We assumed that membrane properties in C. elegans are similar, and used an average of the four measurements: Ri = 180 ficm, and Rm = 150 kficm 2. We assumed a specific membrane ca-pacitance of 1 pF/ cm2, a standard value for a lipid bilayer (Rail, 1989). These membrane properties were adapted to C. elegans anatomy by using the estimated surface area of each cell (see Table 2.2). Each neuron's branching morphology is given in Wood (1988) and White et al. (1986). This, together with measurements of electron micrographs in White 14 Process Length Cell primary secondary surface area C m Rm v(0 Vb (mm) (mm) (10-6cm2) (pF) (Gfi) A L M 0.50 0.03 9.1 9.1 16 0.89 PLM 0.48 0.06 9.1 9.1 16 0.90 AVM 0.24 0.03 5.0 5.0 30 0.97 PVM 0.50 — 8.7 8.7 17 0.89 LUA 0.10 1.4 1.4 107 0.99 PVD 0.74 0.22 16 16 9.4 0.78 PVC 0.96 — 16 16 9.4 0.68 AVA 0.93 - 15 15 10 0.69 AVB 0.86 - 14 14 11 0.73 AVD 0.86 - 14 14 11 0.73 DVA 0.91 - 15 15 10 0.70 Table 2.2: Summary of neuron characteristics. Branching morphology and process length were taken from Wood (1988) and White et al. (1986), assuming a stan-dard worm length of 1 mm. An average process diameter of 0.5 fxm was obtained from measurements of electron micrographs in White et al. (1976) and White et al. (1986). An average soma diameter of 5.0 pm was measured from camera lucida drawings in Wood (1988). V(l)/V0 is the attenuation of a voltage clamp VQ along the full length of the primary process according to a sealed-end cable equation (Rail, 1989), and gives an indication of a cell's isopotentiality. 15 et al. (1986) and White et al. (1976) was used to determine average process lengths and diameters (see Figure 2.2 for an example). Figure 2.2: Example of neuron process length and diameters. This figure depicts a schematic of the AVA cell and electron micrographs taken at different cross sec-tions (labelled e,b,d c). The perimeter of a process was measured at each cross section, corrected for 10% shrinkage due to fixation, and the diameter of cylinder with equivalent perimeter was computed. (Adapted from White et al., 1986). Processes were assumed to be cylindrical and somas were assumed to be spherical. Process diameters varied from 0.2 to 1.0 pm, and an average value of d = 0.5 jira was used. Process lengths were taken from diagrams in Wood (1988) assuming a stan-dard worm length of 1 mm. Soma diameters were taken from camera lucida draw-ings in Wood (1988). Soma diameters varied from 2 pm to 10 fim, but since the soma contributed only a small fraction of the total surface area, we used an average soma diameter of 5 pm. From these data, a total membrane surface area for each cell was computed and the resulting total membrane capacitances and resistances 16 for the entire cell was derived (see Table 2.2). For simplicity we assumed that cells were isopotential. Because the length constant (Rail, 1989) — given by dRm 1 1N •i 1mm, (2.1) where d is the process diameter — was generally longer than the process (data not shown), this isopotential assumption was reasonable (also see Table 2.2). Recent electrophysiological studies in C. elegans by Goodman et al. (1998) support these basic assumptions. They found that the isopotential assumption holds in an identified C. elegans neuron ASER, as well as many other unidentified C. ele-gans neurons. However, several of the interneurons used in our model (AVA, AVB, PVC and AVD) have processes that are significantly longer than that of ASER and no direct recordings have yet been made in these cells to test the isopotential as-sumption. Goodman et al. (1998) also report that the neurons they studied have electrotonic properties that would enable passive signal transmission without spik-ing and that they failed to elicit classical Na+action potentials from all neurons they studied. However, they did not preclude the possibility of Ca2+-dependent spiking resulting from a regenerative calcium current they found in ASER. Thus, a neuron's membrane potential, V, is governed by the usual single-compartment membrane equation (Segev, Fleshman and Burke, 1989): ^m~lZ ~ TT~(BLEAK — V) + 2)2 ^SYN + IINJ, (2.2) at Km where CM is the total membrane capacitance for the cell, Rm is the total membrane leakage resistance for the cell, ELEAK is the equilibrium potential of the cell mem-brane's leakage channels, ISYN is the current attributable to a synaptic input, and IINJ 17 is any injected current. A value of -35 mV was used for E L E A K in these cells (R. E. Davis, personal communication). For the direction of current flow, this document adopts the engineering con-vention that is used in other computational neuroscience works (e.g. Bower and Beeman, 1995): positive current flows in the direction of positive charge. Note that this is in the reverse direction from the convention used by electrophysiologists. 2.4 Gap Junctions The anatomical reconstruction of the nematode nervous system allowed the identification of both electrical and chemical synapses (White et al., 1986). Gap junctions were modelled as ohmic resistances where current flowing into cell i from cell j is given by k = h ( v i - v i ) (2-3) where c/ij is the total conductance of the gap junction area. Niebur (1988) used a specific conductance of 1 S/cm2 (Bennett, 1972) for a patch of membrane area, and used unpublished micrographs to determine the area of each gap junction. He reported that areas of gap junction contact ranged from 0.2 to 2 ^ m long and were 0.5 pm wide (Niebur, 1988). In our model, a standard gap junction length of 1 pm was assumed, with a resulting conductance of 5 nS for all gap junctions. In some experiments, this value was increased or decreased by a factor of 10 to test the sensitivity of the model's predictions to the precise value of the conductance used. 18 2.5 Synapses Synaptic classes consisted of a number of individual synaptic contacts. The number of contacts in each class was extracted from an anatomical database of C. el-egans synaptic connectivity (Achacoso and Yamamoto, 1992). The identification of chemical synapses from the anatomical reconstruction was done by identifying presynaptic specializations and inferring postsynaptic partners based on proximity; no postsynaptic specializations were evident in the electron microscope reconstruc-tion (White et al., 1986). Any error associated with this technique would tend to overestimate the number of chemical synapses in the organism, but these errors would not appear with high frequency, and would therefore not have a large impact on the response of the circuit to stimulation. Each modelled synapse represented a class of synaptic contacts with total synaptic conductance — the "weight" — given by the product of the number of individual contacts within the class and the individual synaptic conductance. The synapse model used was based on the graded synapse model used by Lockery and Sejnowski (1992) in the leech local bending circuit. However, it was extended to explicitly include the synaptic reversal potential as well as the conduc-tance. Post-synaptic current was attributable to gated channels in the post-synaptic membrane with inward current given by I = 9(t)(EsYN-VP0ST), (2.4) where g(t) is the synaptic conductance of the postsynaptic membrane (which is related to neurotransmitter release which in turn depends on presynaptic potential), ESYN is the reversal potential for the synaptic conductance, which was assumed 19 to be constant, and VPosr is the postsynaptic membrane potential. For excitatory synapses, a reversal potential of 0 mV was used, and for inhibitory synapses -45 mV was used (from Ascaris data; R. E. Davis, personal communication). For simplicity, it was assumed that all synapses made by a given presynap-tic cell were of the same polarity and class. This is a version of Dale's Principle (Osborne, 1983) which, although it is not always true, is often used in mathemati-cal neurosciences to simplify models (Hoppensteadt and Izhikevich, 1997). Com-putationally, this reduced the number of optimized parameters — each of which possessed two possible values — to the number of neurons in the modelled circuit. It was further assumed that all modelled synapses functioned as fast ligand-gated channels. It was possible that some anatomically defined synapses were modula-tory and acted via slow second-messenger systems, or that synaptic function was altered by the milieu interieur (Harris-Warrick et al., 1992); however, we assumed that these modulatory effects did not affect a single tap withdrawal response. In the leech local bending reflex, Lockery and Sejnowski (1992) observed multiple time courses in some of the postsynaptic responses and to model this, they used a fast (10 ms) and a slow (1500 ms) decaying membrane current, each gov-erned by its own first order equation. Preliminary versions of this model used a dynamic synaptic model with a fast (10 ms) synaptic time constant, but no sig-nificant differences were noted in the results of circuits containing these synapses and simulations, which used an instantaneous synapse. To reduce the complexity of the model, we therefore used a synaptic conductance that depended only on the 20 presynaptic membrane potential: g{t) = goc(vPRE), (2.5) where represents the steady-state post-synaptic conductance in response to a presynaptic voltage. 2.6 Synaptic Activation No direct recordings have yet been made in C. elegans to determine prop-erties of synaptic activation. In recordings made from Ascaris commissural mo-torneurons, Davis and Stretton (1989b) demonstrated that synaptic transmission is graded and transmitter is released tonically between both excitatory and inhibitory motorneurons and postsynaptic muscle and motorneurons. They found that changes in postsynaptic potential were related to presynaptic depolarizing current by a sig-moidally shaped curve and that the presynaptic resting potential lies approximately in the middle of the voltage-sensitive range of synaptic transmission. Dynamic network simulations based on graded synaptic transmission have been described previously (Lockery, Nowlan and Sejnowski, 1992; De Schutter, Angstadt and Calabrese, 1993). We assumed that synaptic activation and transmis-sion in C. elegans was similar to Ascaris, namely that it is graded and sigmoidally shaped with presynaptic potential, and is tonically active with the steady-state po-tential in the middle of the voltage sensitive range. Accordingly, we used a sig-moidal function to model the steady-state post-synaptic membrane conductance: 9OO{VPRE) -9 (2.6) 21 where g is the maximal post-synaptic membrane conductance for the synapse and Vss is the presynaptic cell's in-circuit steady-state potential, and ERANGE is the presy-naptic voltage range over which the synapse activated. Note that because of tonic synaptic input, the in-circuit steady-state potential of a cell must be determined from the fixed point of the entire system of equations governing the circuit and this was computed prior to each run. We used a value of tf = 21n(^) = -4.3944 (2.7) so that the conductance changes from 10% to 90% of its maximal value over a presynaptic voltage range of ERANGE. Note that because synapses were tonically active, a cell's steady-state potential was not defined solely by its membrane leakage reversal potential, but rather was determined from the steady-state solution of the entire system of equations governing the circuit. This was computed before each run in the following way. 2.7 Steady-state Potential The assumption that tonically active synapses were active in the middle of their voltage sensitive range at the steady-state potential implied that the postsy-naptic conductance g(t) was one-half its maximal value g when the circuit was at steady-state. Let Vi denote the membrane potential for neuron i and similarly for other quantities pertaining to neuron i (see Equation 2.2). Let 7^ denote the ligand-gated 22 synaptic current flowing into neuron i resulting from neurotransmitter release from neuron j across a single synapse, let Esmij denote the synaptic reversal potential for synaptic current 1^, and let n t J denote the total number of synaptic connections from neuron j to neuron i. Similarly, let 1^ denote current flow across a single gap junction where positive current is in the direction from neuron j to neuron i and hij denote the total number of gap junctions between neuron j and neuron i. Finally, let IINJi denote the external current flow into cell i (caused by either sensory stimulation or external current injection). Then the entire system is given by: dVi N RmiCmi—jj- = EiEAKi —Vi + Rmi + Iij) + RmJlNJi (2.8) at j=1 Iij = fujgijiVj -Vi) (2.9) hj = riijgij(EsYNij ~ Vi) (2.10) d9ij _ 9ooM)-9ij (2 11) dt Tij 900M) = (2.12) 1 + e  i j ^RANGEij where N is the number of neurons in the circuit and is the synaptic time constant. At steady-state, Vi — Vss{, and and external inputs IINJi are zero. Synaptic conductances are tonic and at their half-activation at steady-state, so that g0oii(VSSj)=gij/2. (2.13) After algebraic manipulation, this yields a system of linear equations that can be solved using standard Gaussian elimination to find (Press et al., 1988): VSS = A - J b (2.14) 23 where is the ith row and jth column of matrix A and is given by Aij = -Rmirlijgij, i ^ j , (2.15) N An = 1 +Rmi^2(nijgij+ riijgij/2), (2.16) 3=1 and b is a vector is given by N ' bi = ELEAKI + RM. ^2 EsYNijnijgij/2. (2.17) j=i The computed steady-state potential of a cell varied within a physiological range of -47 mV to 0 mV, with a mean of -24 mV and a standard deviation of 13 mV. This did not vary appreciably from cell to cell, but rather depended on the circuit's polarity configuration and ablation condition (i.e. which cells were removed from the circuit). As an aside, the system (2.14) can also be inverted to explicitly give ELEAKI in terms of : N N ELEAK, = VsSi-Rmi^nijgijfeiEsn/ij-VssJ-RmiY^nijfa (2-18) 3=1 3=1 2.8 Synaptic Parameters To determine values for ERANGE and g, the synapse model was fitted to pub-lished measurements (Davis and Stretton, 1989b) of Ascaris muscle cell postsynap-tic response to presynaptic current injection as detailed below. Thus, values for g and ERANGE for particular Ascaris synapses were found. We assumed that C. elegans synapses activated over voltage ranges similar to As-caris synapses. However, the maximal synaptic conductance g needed to be adapted 24 to C. elegans. We assumed that g represents the product of a synaptic conductance per unit area and a synaptic area. In the case of a synapse mediated by a single pop-ulation of ion channels, g would be equivalent to the single-channel conductance times the total number of available channels. To adapt the g value from Ascaris to C. elegans, we assumed C. elegans synapses had similar unit-area conductances and accordingly scaled the Ascaris g by a factor to account for the presumed differ-ence in synaptic areas. We assumed the total synaptic area between two cells was proportional to the length of the process which we estimated by the ratio of body lengths — approximately 1/250. As this represents only a gross approximation, the value of g used in these studies was varied over three orders of magnitude in different experiments (see Wicks, Roehrig and Rankin, 1996). 2.8.1 Modelling Ascaris Monosynaptic Response Data from Davis and Stretton (1989b, their Figs. 13 and 14) show the post-synaptic response of an Ascaris dorsal muscle cell (DM) to current injected into a presynaptic excitatory motorneuron, DEI, and the response of a ventral muscle cell (VM) to current injected into a presynaptic inhibitory motorneuron, VI. Both of these response profiles were sigmoidal in shape, were centred approximately at the resting potential, and had asymptotic post-synaptic responses at the extremes of positive and negative pre-synaptic current injection. Therefore, the sigmoidal tonic synapse model presented here was well-suited to modelling these synaptic responses. 25 Davis and Stretton (1989a; 1989b) placed a recording electrode in a muscle cell within the output zone of the motorneuron, and an injecting electrode at the ventral end of a commissural process leading to the synapse. The measured input resistance of the motorneuron was used to obtain the resulting membrane potential at the point of injection, and an infinite cable model was used to determine the membrane potential at the presynaptic site. Because the recordings that were used to determine input resistance were made at the same ventral end of the commissure as was injected for the synaptic response measurements (R. E. Davis, personal communication) and because the input resistance was approximately constant over the relevant range of injected cur-rent (Davis and Stretton, 1989a), it is possible to directly use the measured input resistance to determine the membrane potential at the point of current injection. Davis and Stretton (1989a) determined the motorneuron cable properties by fitting their measurements along the commissure to an infinite cable model (Rail, 1989), and found that the length constant was unusually high (A « 8mm) — roughly the same magnitude as the length of the process they were measuring. With such a large length constant, it is possible that the cable's branching morphol-ogy and sealed ends play a significant role in determining these cable properties (Rail, 1977), suggesting that a sealed-end cable model might be more appropri-ate. However, for consistency we used an infinite cable model with cable constants as determined by Davis and Stretton (1989a) to reproduce their measured voltage response along the commissure. Specifically, the presynaptic depolarization in response to an injected current 26 is given by AVPRE = IMRPREe-Llx (2.19) where L is the distance from the point of current injection to the synapse, and RPRE is the input resistance at the point of injection. For the DEI - D M synapse, the distance L was 5 - 8 mm, and for the VI-VM synapse, it was 0.5 - 2.5 mm (R. E. Davis, personal communication). We used the mean of 7 mm and 1 mm, respectively. The input resistances for DEI and VI were reported to be 6 Mfi and 17 Mft respectively (Davis and Stretton, 1989a). According to the sigmoidal tonic synapse model, the steady-state plateau response of the post-synaptic muscle is expressed as: Vposr = ELEAK + RPOST n 9oo ( A VPRE) (ESYN — VPOST) , (2.20) where E L E A K and E$YN pertain to the post-synaptic muscle cell, RPOST is the post-synaptic cell's input resistance, goo(AVPRE) is the steady-state synaptic conduc-tance, and n is the number of synapses between the motorneuron and the muscle cell. Note that here, the presynaptic potential is written as AVPRE and is taken rel-ative to the presynaptic cell's steady-state potential since the data points taken for Figure 2.3 and Figure 2.4 were taken relative to steady-state and not to any absolute voltage reference (Davis and Stretton, 1989b). The input resistance of ventral and dorsal muscle cells was measured to be 0.18 - 0.50 MfL with a mean of 0.3 Mfi(R. E. Davis, personal communication). However, it is worth noting that since muscle cells exhibit spiking behaviour, this input resistance may not be truly constant. We used ELEAK = —35 mV and ESYN = 0 mV for excitatory and ESYN = —45 mV for inhibitory reversal potentials (Davis, 27 personal communication). A light microscope study of dye-injected muscle cells in Ascaris suggested that the DEI motorneurons make 5-10 synapses to each dorsal muscle cell and the VI motorneurons make 8-16 synapses to each ventral muscle cell (J. Donmoyer, personal communication). Equation (2.20) can be arranged to give VPOsr explicitly in terms of VPRE: , , ELEAK + ESYN RpOST H goo (AVPRE) /<-><-> 1 \ yposr = T T ~ D TWT7—\ U . z t ; 1 + RPOST n goo{AVPRE) where 9oo(AVPRE) = j L — . (2.22) 1 - | - e  ERANGE The change in postsynaptic potential is therefore given by AVPosr = Vposr — Vss, (2.23) where Vss is the steady-state potential of the postsynaptic cell under unstimulated tonic synaptic input, and is given by T r EiEAK + ESYN RpOST^g/2 /^O^N vss = n , p ZTJT: • (2.14) 1 + Rposrng/2 Equations (2.19) and (2.21) — (2.24) define a non-linear function for post-synaptic membrane potential in terms of presynaptic injected current and contains two unknown parameters, g and ERANGE- Levenberg-Marquardt's method (Press et al., 1988) was used to fit this function to the data from Davis and Stretton (1989b). We obtained good results to the fit for the VI-VM synapse (see Figure 2.3). This fit included the reversal potential ESYN as a fit parameter; this improved the fit substantially without significantly changing the reversal potential (-48 mV as 28 Figure 2.3: Fit for the VI-VM synapse. The diamonds indicate individual measure-ments made of the change in postsynaptic membrane potential in a muscle cell V M in response to a current injected into a presynaptic inhibitory motorneuron VI (ex-perimental data taken from Davis and Stretton (1989b), their Fig. 14). The curve is the fit by equations (2.19) — (2.24) produced by Levenberg-Marquardt's method, with results of g = 150 nS, ERANGE = 52 mV, and ESYN = —48 mV. 29 opposed to -45 mV). The fit results were g = 150 nS, ERANGE = 52 mV, ESYN — —48 mV and were stable under various initial conditions. The DEI-DM fit was less accurate since the steady-state conductance was not precisely a symmetric sigmoid. Therefore, to find approximate ranges for ERANGE and g, a number of parameter values were explored manually to fit each branch of the sigmoid separately, and we found that ERANGE ranged from 13-20 mV, and g ranged from 50-150 nS (see Figure 2.4). DE1-DM 6 I 1 1 1 1 1 1 1 1 r -8 I i i i i i i i i i I -10 -8 -6 -4 -2 0 2 4 6 8 10 linj (nA) Figure 2.4: Fit for the DE1-DM synapse. The diamonds indicate individual mea-surements made of the change in postsynaptic membrane potential in a muscle cell DM in response to a current injected into a presynaptic excitatory motorneuron DEI (experimental data taken from Davis and Stretton (1989b), their Fig. 13). Be-cause the data do not form a perfect sigmoid, a good fit was not obtained to both branches simultaneously. The solid curve shows a fit to the upper branch (g = 150 nS; ERANGE =13 mV; ESYN = 0 mV), and the dashed curve shows a fit to the lower branch (g = 50 nS; ERANGE = 20 mV; ESYN = 0 mV). In both cases, the curves were fit by hand. 30 To adapt these results to C. elegans we took a rough average of the two ERANGE estimates to get an activation range of -35 mV, and scaled the g estimate from the VI-VM synapse by the 1/250 ratio of body lengths to obtain a maximal conductance of 0.6 nS for an individual C. elegans synapse. 2.9 The Gearbox This model does not explicitly incorporate nematode locomotion; these is-sues have been dealt with adequately elsewhere (Niebur and Erdos, 1991; Niebur and Erdos, 1993). Rather, this report concentrates specifically on sensorimotor in-tegration. However, because a behavioural variable was used to optimize the mod-elled output, it was necessary to rigorously define the relationship between the an-imal's locomotion and activation of the circuitry that controls that behaviour. This issue was addressed with simple assumptions, which were consistent with work on the modelling of nematode locomotion (Niebur and Erdos, 1991; Niebur and Erdos, 1993; Stretton et al., 1992) and current theories of tap withdrawal circuit function (Chalfie et al., 1985; Wicks and Rankin, 1995). The output of the tap with-drawal circuit was assumed to control locomotory behaviour primarily through the action of the interneurons AVB and AVA. These two interneurons make electrical connections with motorneurons all along the ventral cord of the worm. The AVA interneurons make gap junctions with the motorneurons AS, VA, and DA, which are presumed to excite backward locomotion; the AVB interneurons form gap junctions with the motorneurons VB and DB, which are presumed to excite forward locomo-tion. Ablation of these cells almost completely destroys an animal's ability to move 31 forward (in the case of AVB ablations) or backward (in the case of AVA ablations) (Chalfie et al., 1985; Wicks and Rankin, 1995). Thus, it was simply assumed that the degree to which an animal reversed was proportional to the depolarization of the AVA interneuron and inversely proportional to the depolarization of the AVB interneuron. Forward locomotion in response to tap was also proportional to this value; a lower propensity to reverse was equivalent to a higher propensity to ac-celerate. The exact nature of this proportionality was not defined because in vivo it will be modulated by a number of neural, hydrostatic, and physical forces that are beyond the scope of this endeavour. The gearbox, i.e., the transformation equa-tion that was used to convert depolarization of AVA and AVB into behaviour, was simply: Propensity to Reverse oc J(VAVB — VAVA) dt. (2.25) The integration was calculated from the time of the tap stimulation until either the end of the simulation or until the integrand changed sign. Additionally, the test for a change of integrand sign was suppressed for a grace period of 100 ms to allow for initial transients after the tap. The tap stimulus was modelled as a phasic depo-larization of the sensory neurons PLM, A L M and AVM which have been shown to mediate the response to tap in the intact animal (Wicks and Rankin, 1995). One consequence of the gearbox assumption is that, because of uncertainty regarding the exact nature of the proportionality between the output of the AVA and AVB interneurons and the magnitude of the evoked behaviour, comparisons of model data and empirical data must be limited to relative changes in response magnitude. Thus, such comparisons were made between data profiles that had been 32 normalized about the mean of that polarity configuration's response level (i.e. the value of the integral (2.25) for the model data, and the living animal's reversal magnitude for the empirical data. See Wicks, Roehrig and Rankin, 1996) This measure detected changes in the levels of responding to tap produced by an ablation series, without being sensitive to the absolute response magnitude of a particular circuit configuration — information which in any case is meaningless in the context of the gearbox assumption. 2.10 Results and Conclusions This model was successfully used (Wicks, Roehrig and Rankin, 1996) to predict the functional polarities of synaptic connections in the tap withdrawal cir-cuit in the following manner. Because the synaptic weights were determined by anatomical data and other model parameters were determined from physiological criteria, the only free parameters in the model were the synaptic polarities. The out-put of the model was then tested against the behaviour of the living animal under various conditions of degradation involving the removal of one or more cells from the circuit (using laser ablation to kill the cell in the animal). For each ablation con-dition, all possible synaptic polarity configurations of the model were exhaustively enumerated and tested to find the best overall fit to the behavioural data. The model provided statistically significant predictions for 13 of the 15 neuron classes and ac-curately reproduced the response of the animal to ablations (see Figure 2.5). For full details on the polarity determination strategy and results, the reader is referred to Wicks, Roehrig and Rankin (1996). 33 ABLATION DATA CON PLM PVC PVD AVM ALM ALMAVM MODEL DATA z o CON PLM PVC PVD AVM ALM ALMAVM Figure 2.5: Comparison of model data to behavioural data. The top figure shows the reversal response of the animal, normalized to the control (CON) response. The bottom figure shows the response of the model to the same set of ablations when the predicted synaptic polarities are used. (Adapted from Wicks, Roehrig and Rankin, 1996). 34 A novel contribution of this model is the tonic and graded synaptic model it uses and the resulting distinction that it makes between a cell's resting membrane potential and its in-circuit equilibrium potential. Graded (non-firing) synaptic mod-els have been proposed previously (Lockery et al., 1992; De Schutter et al., 1993), but our assumption that synapses were tonically active forced us to consider whether they were tonically active in an isolated cell or while in a circuit. In the latter case, the half-activation threshold of the cell is its in-circuit equilibrium potential which differs from its leakage membrane potential due to the tonic synaptic currents flow-ing in the circuit. The preceding derivation contains a treatment of the equilibrium vs. resting potential and formulas to convert from one to the other. For the purposes of this dissertation, the interesting consequence of this model is that the output of the model (the gearbox), while successfully predict-ing the magnitude of the behavioural response in the animal, did not follow the same time course as the animal's behaviour. Following a transient tap stimulus last-ing from 100-600 ms, the nematode continues to swim backwards for up to several seconds before resuming forward locomotion. The gearbox output of the model, on the other hand, decays quickly after the termination of the stimulus — with a time constant equivalent to the average cell membrane time constant, approximately 150 ms. To account for this discrepancy, a new model that uses network dynamics to maintain the reversal was proposed, and this is the subject of the next chapter. 35 Chapter 3 Dymods: An Approach to Modularizing Dynamical Neuronal Structures The physiological model of the tap withdrawal circuit we developed in Chapter 2 successfully predicted the behavioural response, but it did not explain how the reversal locomotion was maintained for several seconds after the comple-tion of the transient tap stimulus which lasts for only a few hundred milliseconds. The duration of a reversal response is much longer than both the stimulus and the characteristic time constants of the neurons that participate in the response (Wicks, Roehrig and Rankin, 1996; Goodman, Hall, Avery and Lockery, 1998). Moreover, the crispness of the change from forward locomotion to reverse and back again to forward suggests that it is an active transition between two states rather than a decay of an impulse that is responsible for maintaining and terminating the reversal. 36 One possibility is that the maintenance of a reversal could be the result of network dynamics and a bifurcation between two quasi-stable states governing for-ward and reverse locomotion. To test this hypothesis, physiologically-based models of C. elegans cells were used to engineer a neuronal circuit to account for the rever-sal maintenance response. Because the goal was to see if network dynamics could account for the behaviour and because so little is known about cellular activity in C. elegans the design was accomplished without the connectivity constraints of the anatomical data. The way that the resulting circuit might fit into the anatomical map is addressed in the discussion. 3.1 Methods and Results The initial conceptual design of this circuit consisted of a pair of cells to drive forward and reverse locomotion. The tap was assumed to cause the forward drive cell to inactivate, the reverse drive cell to activate, and a charging circuit to begin a charging cycle. When the charging circuit reached a threshold, it was as-sumed to cause the reversal cell to inactivate and the forward cell to reactivate. Because there is some evidence that an oscillatory signal is needed to drive loco-motion (Niebur and Erdos, 1991), the circuit was assumed to exhibit oscillatory activity. A current input pulse lasting 100 ms was used as an approximation to the tap stimulus (Wicks, Roehrig and Rankin, 1996). The resulting neuronal circuit model for maintaining reversals is a dynam-ical system and it was constructed in a modular way from three dymods, each themselves a dynamical system: a bistable switch dymod,-a.bistable oscillator dy-37 mod and a charger dymod. These dymods were designed independently and in-terconnected to form the complete circuit. The dymods were designed using the XPP phase plane analysis software for non-linear systems of differential equations (Ermentrout, 1998). Al l graphs were produced from the output of XPP computa-tions. 3.1.1 The Cell Model C. elegans neurons are simple. Although electrophysiology in C. elegans is in its infancy, current results suggest that signalling in C. elegans neurons is accom-plished passively without spiking (Goodman et al., 1998). Direct recordings from some neurons (Goodman et al., 1998) and estimations based on data from a related nematode species, Ascaris (Wicks et al., 1996; Davis and Stretton, 1989a) suggest that neurons are nearly isopotential and can be approximated by a single mem-brane equation, equivalent to a passive RC circuit (Wicks et al., 1996; Koch and Segev, 1989). Synapses in C. elegans are also simple: evidence from Ascaris (Davis and Stretton, 1989b) suggests that neurotransmitter is released tonically (even when the cell is at "rest"), so that positive or negative changes in the cell's activity (i.e. its membrane potential) result in proportional changes in synaptic transmitter out-put. In the cell model, the synaptic activation function describes how much neu-rotransmitter is released at a given presynaptic cell membrane potential. Based on data from Ascaris (Davis and Stretton, 1989b), this function was assumed to be sigmoidally shaped: it is approximately linear at the cell's resting potential and saturates smoothly at both extremes. As in our previous study (Wicks, Roehrig 38 and Rankin, 1996), a simple equivalent channel model was used for synaptic input: current flows into the postsynaptic cell with a driving force equal to the difference between the post-synaptic cell's membrane potential and an effective channel re-versal potential. This reversal potential determines the synaptic polarity (whether it is excitatory or inhibitory). For full details about the physiological derivation of the model and its assumptions, see Wicks, Roehrig and Rankin (1996) or Chapter 2. Figure 3.1: A coupled pair of cells. The design of the dynamical system for maintaining reversals was begun by examining the dynamics of a pair of two reciprocally connected cells (Figure 3.1). This system has two types of behaviour. The first is simple: each cell's membrane potential tends toward a stable steady-state voltage that is dependent on the synaptic efficacies (weights). The second is more interesting: a pair of cells can operate as a switch that can toggle between two different stable states in response to an external transient input. The voltage response of a pair of coupled cells is described by the following system: m -V1 + J12(V2) ( 2 W - V i ) dt T\ dV2 -V2 + MV,) (ESYN21-V2) + IINJI (3-1) + IINJ2 (3.2) dt T2 where Vi is the membrane potential of cell i (i = 1 or 2) taken with respect to the cell's leakage (resting) potential, r,- is the membrane time constant for cell i, ESYNIJ is the synaptic reversal potential for connections into cell i from cell j, taken with 39 respect to the resting potential of cell i (the postsynaptic cell). The sign of Esmij determines whether the synapse is excitatory (positive) or inhibitory (negative). The derivation of this model is presented in Wicks, Roehrig and Rankin (1996) and Chapter 2, but note that voltage quantities here are expressed relative to the cell's leakage potential, rather than as the absolute quantities used in Chapter 2. Jij (Vj) is the synaptic input to cell i from cell j and is given by Jij{V-) = wn s(Vj~EACTj), (3.3) GRANGE] where EACTJ is the synaptic half-activation potential for cell j , and E'RANGEJ is the voltage range over which cell j activates. The weight coefficient uiij is the dimensionless synaptic coupling for con-nections into cell i from cell j . In physiological terms, Wij is the ratio of the fully-activated postsynaptic conductance (gsYN) to the postsynaptic cell's leakage conductance (C/LEAK\ multiplied by total number of synapses. The sigmoidal synaptic activation proportion is given by s(x) = l+\k'x, where K = -21n(£f) = -4.3944 so that the synapse activates from 10% to 90% over a voltage range of ERANGE-The external input to cell i is expressed as Imji which is the change in the cell's membrane potential induced by an injected current I across the cell's input resistance, Rmpw, and is expressed in mV/s: IINJ = I x RINPUT, where the input resistance was taken to be a constant 10 GQ for the short duration of the pulse input used in the experiments (Wicks et al., 1996). 40 3.1.2 The Switch Dymod The possible behaviours of this system can be deduced by studying the null-clines of the system in the (Vi, V2) phase-space. (The nullcline is the set of points where the membrane potential of the cell does not change.) At an intersection of the two nullclines, both cells are unchanging so this defines a steady-state or fixed point of the system (Strogatz, 1994). 100 50 h v2 0 -50 -100 1 v1 = o 1 r 1 v2 = o \ • -100 -50 50 100 Figure 3.2: Nullclines for a coupled pair of cells. In this case, cell Vi makes an excitatory connection to cell V2 (ESYN2I > 0). and cell V2 makes an inhibitory connection to cell V\ (ESYN\2 < 0). The nullclines of the system (3.1) and (3.2) are given by ^ = 0 and ^ = 0, and can be written as: y _ W ) J2i(Vi) F yl - T-—rTT7TEsYNi2,  v2 - 7TT1T7T^«V21 (3.4) l + JM) d " V i Z ' ' I + J21W In the (Vi, V2) phase plane, these nullclines are sigmoidal in shape and have asymptotes parallel to their respective axes (see Figure 3.2). In the simple case, 41 Vj ( m V ) Figure 3.3: Fixed points for an inhibitory switch. In this case, both cells make in-hibitory connections with one another (E$YN < 0), and the synaptic weights and activation ranges have been adjusted so that the nullclines V\ = 0 and V2 — 0 inter-sect at three fixed points (circles). Stable and unstable manifolds for the hyperbolic middle fixed point are shown in dashed lines. See text for parameter values. the two nullclines intersect at a single point which defines the steady-state point for the system. That this point is stable can be deduced from the sign of ^ in each quadrant. It is also possible for the two sigmoids to intersect at more than one point and this underlies the second, more interesting behaviour of the system. The sigmoids can have multiple intersections in the third quadrant when both connections are inhibitory (Figure 3.3), or in the first quadrant when both connec-tions are excitatory. In both cases, there are three intersections corresponding to three equilibrium states of the system. The outermost two fixed points are stable, and the middle one is an unstable hyperbolic point. The excitatory case has stable points when the cells are either both activated or both inactive. The inhibitory case has stable points when only one is active and this is more suitable for the design 42 since only one of the forward and reverse drive cells can be active at a time. Figure 3.3 also shows the stable and unstable manifolds for the middle hy-perbolic point. Trajectories that start precisely on the stable manifold will tend toward the middle hyperbolic point. All others will tend toward one of the outer sta-ble fixed points, so that the stable manifold is a separatrix which separates (Vi, V2) space into two basins of attraction for their respective outer stable fixed point. Fig-ure 3.4 and Figure 3.5 show how this circuit can operate as a neural switch. 5 0 -5 -10 V2 -15 (mV) -20 -25 -30 -35 -35 -30 -25 -20 -15 -10 -5 0 5 Vj (mV) Figure 3.4: Inhibitory switch turning on. The trajectory starts at stable fixed point A with cell 1 turned on (0 mV), and cell 2 off (-32 mV). At t = 200 ms, an inhibitory current pulse of -2.5 pA is injected into cell 1 for 100 ms, causing the trajectory to cross the threshold (Figure 3.4), after which the system settles into its other stable state. This behaviour is not singular and occurs within the physiological ranges of parameter values (see Wicks, Roehrig and Rankin, 1996). Parameter values for this example are shown in Table 3.1. In subsequent design phases, the cells were taken to have these default parameter values, unless the design necessitated a change. 43 'mj. 50 mV 200 ms Figure 3.5: Simulated recording traces for an inhibitory switch. The trajectory starts at a stable fixed point (Figure 3.4) with cell 1 turned on (0 mV), and cell 2 off (-32 mV). At t = 200 ms, an inhibitory current pulse of -2.5 pA is injected into cell 1 for 100 ms, causing the trajectory to cross the threshold (Figure 3.4), after which the system settles into its other stable state. At 700 ms, an inhibitory current pulse is injected into cell 2 and the system switches back to its original state. 44 Note that EAcr = 0 mV corresponds to the assumption that synapses are tonically active at their resting potential. Model Parameters Value r 75 ms ERANGE 20 mV Wij 25 EACT 0 mV ESYN -35 mV IINJ -2.5 pA x 10 GQ = 25 mV/s Table 3.1: Default model parameters used for all cell models unless otherwise noted. The switch is a dymod that forms the basis for the reversal maintenance circuit. It has two outputs: Vi which drives forward locomotion, and V2 which drives reverse locomotion. It has two hyperpolarizing pulse inputs IINJi and ItNJ2 which are used to toggle the switch on and off. 3.1.3 The Oscillator Dymod The switch dymod has two stable output states which are suitable for switch-ing between forward and reverse locomotion, but they are static. To satisfy the requirement of an oscillatory output, the next step was to investigate the possible oscillatory behaviour of these model cells. With these model cells, oscillations are not possible with only two cells (al-though see the discussion following). Oscillations can emerge when a third cell is added to the circuit. In three dimensions, the nullclines become null surfaces, and it is quite difficult to visualize the ways that three null surfaces might intersect and the kinds of trajectories that might arise. Optimization techniques such as genetic al-45 gorithms have been used to construct neuronal dynamical systems in the past (Beer and Gallagher, 1992; Yamauchi and Beer, 1994), but because it was difficult to ex-press a suitable objective function for the desired behaviour, a direct engineering approach was taken. The oscillator dymod was built up from the two-cell inhibitory switch by adding an third cell (Figure 3.6). The basic idea was that the third cell would be made to "pull" the trajectory in its opposite direction, but operating with a delay to result in oscillations. Figure 3.6: Oscillator circuit Because the two switch cells already made inhibitory connections with one another, they were assumed to be restricted to making inhibitory connections with the third cell. This is the same application of Dale's Principle as used in Chapter 2 and it was applied with the understanding that it is not always true. Because of this inhibitory connection, cell 3 would therefore become activated as cell 1 turns off, and by making an inhibitory connection to cell 2, it would "pull" cell 2 off in the desired manner. When the third cell was added, this system produced a new stable fixed point instead of oscillations because the activation of cell 3 was not delayed with respect to cell 1. A delay was added by reducing cell 3's activation range to 5 mV, but leaving its half-activation potential the same as the other cells (to preserve the tonic 46 synapse assumption). The increased steepness of the activation curve gave cell 3 an effective activation delay compared to cell 1 (see Figure 3.7). (This "delay" is not a "hard" delay in the sense of AT(t) = t — r, but rather an apparent delay induced by different activation characteristics.) The activation range was chosen to be steep enough so that cell 3's synaptic transmission only occurred when cell 3 was almost maximally activated, resulting in a pulse-like activation pattern (see the s 3 trace in Figure 3.8). Note that the assumption that cells are tonically active at their resting (leakage) potential does not necessarily imply that they are active at their in-circuit steady-state potential. In fact, in the oscillator, cells 2 and 3 never attain their leakage potential (0 mV in the simulated traces). See Section 2.7 or Wicks, Roehrig and Rankin (1996) for a treatment of steady-state potential versus leakage potential. i synaptic activation 0 presynaptic potential Figure 3.7: Delay due to increased steepness of activation curve. Cell 3 has steeper activation characteristics than cell 1. Because both are centred at EACT (i.e. they are tonically active), cell 1 will begin to activate earlier at a lower presynaptic potential than cell 3. Adding this delay did indeed result in oscillations which tended toward a limit cycle (Figure 3.9). As with the switch, an inhibitory current pulse of 2.5 pA into cell 1 was used to start the oscillations and the same pulse into cell 2 was used 47 'uv2 IT J 500 ms 50 mV 75% activation Figure 3.8: Simulated recording traces for an oscillator. The s, curve is the synaptic VJ-EACTJ ERANGE i activation for cell i (0 < s,- < 1) given by Sj = s(^^^1). 48 to stop the oscillations (Figure 3.8). 5 0 -5 -10 -15 ( m V ) - 2 0 -25 -30 -35 -35 -30 -25 -20 -15 -10 -5 0 5 Vj ( m V ) Figure 3.9: Phase portrait of an oscillating switch. This graph shows the projection of the phase space trajectory onto the (Vi, V2) plane. The trajectory starts at the lower right-hand fixed point and settles onto a limit cycle before an inhibitory pulse returns it to its stable state. 3.1.4 The Charger Dymod The next step in constructing the reversal maintenance circuit was to design the circuit to automatically generate the inhibitory pulse into cell 2 that stops the oscillation. This was done using a charger dymod that received input from the oscillator and gradually increased its activity over a few seconds before exceeding a threshold and delivering the inhibitory pulse to cell 2. A single cell did not easily work as a charger, since it needed excitatory input in order to charge to a threshold, and so far the circuit design only had in-hibitory neurons in it. So again, the two-cell inhibitory switch was chosen as the 49 starting point for the charger dymod. The charger dymod was designed separately as an independent module, with a periodic pulse train test input. The input cell that received the pulse train was labeled cell 4 and its partner, cell 5. When used as input to the switch, a periodic pulse train induced forced os-cillations which can be clearly seen in ( V 4 , V 5 ) phase space (Figure 3.10). There are three different qualitative behaviours that can occur, depending on the strength of the pulse input. For small pulse inputs, the switch is driven into oscillations which settle onto a limit cycle around the nullclines near the initial fixed point. For large inputs, the switch is driven first into its other state and then into small stable oscillations near the other fixed point. v5 ( m V ) 5 r—i 1 1 1 1 1 1 1 \ 0 -5 \ • -10 ^ A v .'" -15 \^_^ '' • -20 • -25 • -30 --35 l ' _ - ' 1 • -40 . 1 i 1 1 1 1 1 i 1 -35 -30 -25 -20 -15 -10 -5 0 5 V4 ( m V ) Figure 3.10: Phase portrait of charging behaviour. The input pulse train kicks the trajectory a bit further each cycle until it finally crosses the threshold. (Input strength of 14.7mV/s). Between these extremes is a fairly narrow, continuous range where the forced oscillations slowly creep along the nullclines until the trajectory nears the 50 switching threshold at which point it is driven across and the switch changes state (Figure 3.10). This is the desired charging behaviour. The charging time depends on the input level, and is determined by the number of oscillations that occur before the trajectory crosses the threshold. Figure 3.11 shows the simulated recording traces for the charging behaviour. The synaptic activation of Cell 5 is suppressed until the trajectory crosses the threshold, at which point it becomes activated. Cell 5 was therefore used as the output of the charger dymod to deliver the inhibitory "turn-off' signal to the oscillator. v4 ys s5 500 ms Figure 3.11: Simulated recording traces of charging circuit. The Sj curves are as defined in Figure 3.8. This charging behaviour, although quite sensitive to the input level, is not a singular occurrence and can occur whenever the nullclines have 3 intersections njirLTLnnriTLTinj 50 mV 75% activation 51 to form a switch. However, the appropriate input range for charging is sensitive to the precise configuration of the nullclines, and therefore depends on the other cells' parameters. In the example in Figure 3.10, the charging behaviour occurred over an input range of about 0.1 mV/s. 3.1.5 Dymod Assembly: The Reversal Maintenance Circuit The final stage in the construction of the reversal maintenance circuit was to connect the charger dymod to the oscillating switch dymod to turn it off (Fig-ure 3.12). Figure 3.12: Complete reversal maintenance circuit. Cell 3 was chosen as input to the charger dymod because its synaptic output was pulse-like (see S3 in Figure 3.8). The charging behaviour of the circuit was first tuned with its output disconnected by adjusting the input coupling strength to cell 4 ( W 4 3 ) until the appropriate charging behaviour was manifest. This was done by observing the charging trajectory in (V4, V5) phase space as in the design of the charger. Cell 5 was then connected to Cell 2 to deliver the inhibitory "turn-off" signal, and its coupling strength (W25 ) was gradually turned up until the oscillator turned off. During this process, the input coupling strength to cell 4 needed to be adjusted to compensate for the interactions between the oscillator and the charger and maintain the appropriate charging behaviour. The process was straight-forward 52 and required only a few iterations of adjustments. The simulated recording traces of the complete circuit are shown in Figure 3.13. Note that the shape of the charger traces (V^V^) in Figure 3.13 is more rounded than those of Figure 3.11. This could be because a more rounded pulse ( « 3 in Figure 3.13) was used as charger input (compared to the square pulse of Figure 3.11), but it might also be due to some non-linear interactions between the oscillator and charger dymods. 3.2 Discussion In this chapter a circuit was designed to help explain how the nematode C. elegans makes transitions between forward and reverse locomotory modes and maintains its reversal for a duration much longer than the time constants of the cells controlling the behaviour. The circuit was designed using the novel approach of combining separate dynamical modules ("dymods") governed by non-linear ordi-nary differential equations. In spite of the non-linear interactions between dymods, the assembly was remarkably straight-forward suggesting that a human-engineered approach might not be so difficult as it seems. The system is depicted in Figure 3.14 and consists of three multi-cell dynam-ical modules (dymods): a switch, an oscillator, and a charger. The switch bifurcates between two stable states in response to a tap stimulus: a stable forward state where Vi drives forward locomotion and V2 is off, and a reverse state where V\ is off and V2 drives reverse locomotion. The oscillator was constructed out of the switch by adding an "oscillator helper" cell V3 to "pull" V2 against its trajectory, but with a 53 'mjj Vl FORWARD DRIVE REVERSE DRIVE OSCILLATOR HELPER CHARGER (INPUT) CHARGER (OUTPUT) 50 mV 75% activation 500 ms Figure 3.13: Simulated recording traces for the complete reversal circuit. The curves are as defined in Figure 3.8. Non-default parameter values: ERANGES 5 mV; W2§ = 5; = 30. See Table 3.1 for other parameter values. 54 osc i l l a tor FWD REV Figure 3.14: The dymod view of the reversal maintenance circuit. "soft" delay (as opposed to a "hard" t — r delay) which was induced by adjusting the synaptic activation characteristics of V3. The oscillator's job was to generate an oscillatory locomotion signal for reverse locomotion. (Note that the forward signal is not oscillatory, but it can be made so by connecting it to another oscillator.) The reverse locomotion is turned off by a signal from the charger (V5). The charger is a variant of the switch which gradually approaches its switching threshold in re-sponse to an oscillatory input from V3. When the threshold is reached, it switches state and Vj> delivers its "turn off' signal to end the reversal. The circuit's operation can be seen in Figure 3.13. 3.2.1 The Locus of C. elegans Reversal Dynamics The reversal maintenance circuit constructed here was based on physiolog-ically realistic cell models, but did not take into account the known anatomical connectivity of the nervous system. The function of the two switch cells V\ and 55 V2 bear a striking resemblance to that of two pairs of ventral cord interneurons: Vi corresponds to the bilateral pair AVBL and AVBR which connect to the for-ward motorneuron pool and V2 corresponds to the pair AVAL and AVAR which connect to the reverse motorneuron pool (White et al., 1986; Chalfie et al., 1985). However, neither AVAL nor AVAR make direct reciprocal connections with either AVBL or AVBR (White et al., 1986) which suggests that they do not function as a switch of the form Vi,V2. Another interneuron pair, PVC, makes strong recip-rocal connections with AVA and is involved in the response to tail-touch (Wicks and Rankin, 1995). However, laser ablation of PVC does not affect the dura-tion of spontaneous reversals (Wicks and Rankin, 1997), which suggests that it is not involved in the reversal maintenance dynamics. As no other interneurons from the tap withdrawal circuit make any strong reciprocal connections (Wicks and Rankin, 1995; Chalfie et al., 1985; White et al., 1986), this suggests that if reversal dynamics is governed by a circuit like that of Figure 3.12, the circuit likely does not lie in the interneuronal circuitry and may instead lie downstream in the motorneuron circuitry or may involve other processes than network level dynamics. Mechanical interactions may play a role in reversal dynamics. When a re-versal is initiated, the body goes into a deep bend which is thought to activate a stretch-related current (Tavernakis, Shreffler, Wang and Driscoll, 1997; Corey and Garcia-Anoveros, 1996) which could modulate neuronal function. Niebur and Erdos's (1991) model of C. elegans locomotion circuitry suggests that interactions between the interneurons, muscle cells, body wall and the medium in which the nematode swims all may play a role in determining the locomotion dynamics. It 56 may be possible, however, to rule out an external component in the reversal dynam-ics. When the worm is placed in water, thus removing the mechanical resistance to body movement, it thrashes in an oscillatory manner during periods of forward locomotion and bends into an unmoving hoop during periods of spontaneous re-verse locomotion before resuming thrashing If the duration of this "reversal" period is unchanged in water, it would suggest that the reversal maintenance dynamics is generated entirely internally without interaction with the medium in which it swims. It is also possible that reversal dynamics is governed by more diffuse pro-cesses which modulate synaptic activity such as neuropeptides (Schinkmann and Li , 1992). A slowly increasing neuropeptide concentration might be a more plausi-ble charging mechanism than the charger presented in this chapter which has very sensitive charging regime and would need some kind of precise regulation. In order to reduce the complexity of the model, gap junctions were ignored during the design process. However, there are strong gap junctions connecting the tap withdrawal interneurons to the motorneurons governing locomotion (Chalfie et al., 1985) and they may also play a role in the reversal maintenance dynamics. Further experiments in the animal could narrow down the possibilities for the locus of the reversal dynamics. Chalfie et al. (1985) and Wicks and Rankin (1995) have successfully used laser ablations to determine the functional roles of neurons in the reversal response to mechanosensation. If the charging component of the reversal dynamics was performed by a neuronal circuit like V4 and V5 in Figure 3.12, one would expect that a suitable ablation would destroy the charging behaviour and the reversal would continue without stopping. This has not been the case for the 57 ablations performed on any of the tap withdrawal interneurons (Wicks and Rankin, 1995; Wicks and Rankin, 1997). It may be possible to perform an ablation study of the locomotory motorneurons to find such a cell, but if the charging mechanism is a modulatory peptide rather than a neuronal circuit, no such cell would be found. Modelling efforts in C. elegans are frustrated by the inability to record neu-ronal activity during behaviour. It appears unlikely that the usual electrophysio-logical recording methods will ever be possible in vivo in C. elegans since its ex-ternal cuticle is under hydrostatic pressure and the worm explodes when it is punc-tured. However, new genetic techniques have been developed to engineer C. elegans strains that express green fluorescent proteins (GFP) in specific cells (Chalfie, Tu, Euskirchen, Ward and Prasher, 1994) and it may be possible in the future to obtain activity information from voltage-sensitive fluorescence resonance energy transfer (FRET) imaging (Gonzalez and Tsien, 1995) using fluorescent proteins expressed in the cells of interest. For the quantitative studying of locomotory behaviour, the most interesting cells are AVB and AVA, the command interneurons that play a key role in mediating forward and reverse locomotion, respectively. Knowing the activ-ity patterns of these cells during a reversal — whether they have transient activity, prolonged activity or oscillatory activity — is crucial to further modelling efforts. 3.2.2 Oscillations Niebur and Erdos (1991) found in their computer model that an oscillatory signal was needed to drive the locomotion circuitry and so the model circuit was as-sumed to exhibit oscillatory behaviour. While this is weak evidence of the existence 58 of a neuronal oscillator, other evidence of intrinsic rhythmic signals in the animal (e.g. the defecation cycle (Thomas, 1990), and side-to-side head oscillations) led us to consider how oscillations might arise from these tonic cells. Additionally, in the initial concept of the charger it was considered that a rhythmic pulsed input would allow for a longer charging time and less sensitivity to the input strength than if a continuously increasing input signal was used, but this has not been verified. In the model, oscillations were obtained by adding a third "delay" cell into the switch circuit (Figure 3.1). This delay was accomplished within the constraints of the simple cell model by using a narrow synaptic activation range (Figure 3.7). However, there may be other ways to obtain oscillations. The dynamic processes underlying synaptic transmission could provide a sufficient delay for oscillations. The cell model used here did not include synaptic dynamics but instead considered synapses to act instantaneously. In earlier work with these model cells (Wicks et al., 1996), we found that a simple model of synaptic dynamics that used a first-order relaxation equation and a 10 ms time constant made no appreciable difference to our simulations. However this synaptic model also did not include a hard delay and it may be possible to obtain oscillations with a synaptic model that included hard delays. Beer (1995) has shown that oscillations can arise in a 2-cell circuit when the cells have self-connections. The cell models he used are continuous time recurrent neural network (CTRNN) cells (Funahashi and Nakamura, 1993) and are very sim-ilar to the C. elegans cell model used here, differing only in that CTRNN cells use a synaptic driving force that is assumed to be constant. Although cells in C. elegans 59 are typically unbranched cylindrical processes and are unlikely to form anatomical self-connections (White et al., 1986), Goodman et al. (1998) have recently discov-ered a regenerative calcium current in a C. elegans neuron which could achieve the same effect as a self-connection. A vast amount of work has been done on neuronal oscillations. Some related modelling work in other biological systems include the central pattern generator (CPG) of the lobster stomatogastric ganglion (Marder and Selverston, 1992; Ab-bott, Marder and Hooper, 1991) which uses a bursting "pacemaker" cell to generate rhythmic oscillations, and the Tritonia sea slug locomotory CPG which generates oscillations at a network level (Getting, 1989). Jung, Kiemel and Cohen (1996) have undertaken a bifurcation analysis of the dynamics of a computational model of the lamprey locomotor CPG. Considerable theoretical work has been done on the dynamics of neuronal oscillations (Hoppensteadt and Izhikevich, 1997; Ermentrout and Kopell, 1990; Beer, 1995), and its physiological mechanisms (Skinner, Kopell and Marder, 1994). Campbell and Wang (1998) have analyzed the dynamics of neu-ronal oscillators that use hard delays. Yang and Dillon (1994) have proven using cells similar to the C. elegans model that 2-cell oscillations not possible without self-connections, and Yang (1995) has analyzed a 3-cell network oscillator similar to the one derived here. It is also interesting to note that C. elegans cells may have some generality as model cells. They are almost the same as CTRNN cells which have been shown to be universal dynamics approximators (Funahashi and Nakamura, 1993). Although C. elegans neurons do not exhibit classical spiking behaviour, passive cell models 60 can be used to model spiking systems by considering cell activity to represent the firing frequency (Bialek, Rieke, de Ruyter van Steveninck and Warland, 1991; Jung et al., 1996). However, this should be tempered with evidence that temporal spiking patterns can encode essential information (Hopfield, 1995). 3.2.3 The Forward-Engineering Approach In this chapter a forward-engineering approach was used to design a circuit whose operation resembles the reversal behaviour observed in the worm. It is nat-ural to ask why this artificial system should tell us anything about reversal mecha-nism in the worm or give insight into biological systems in general, especially when there are so many other possible candidate mechanisms. In the worm, this forward-engineered model elucidates the key concepts in the reversal mechanism: the active transition between two locomotory states and a charger that builds to a threshold, and it provides quantitative evidence that this transition could be accomplished by network dynamics alone, even using simple and physiologically plausible tonic cells. Even if such a mechanism is ultimately not found in the worm, this result has a useful generality. The model also pro-vides a quantitative framework for adding new data about cellular and intra-cellular activity in C. elegans as it becomes available, and it suggests an approach using dynamical phase-plane analysis that could be used in subsequent models. Similar forward engineering approaches have been used to model the leech local-bending reflex (Lockery and Sejnowski, 1992), cockroach locomotion and escape (Beer and Chiel, 1993), and chemotaxis in C. elegans (Morse, Ferree and Lockery, 1998). In 61 these studies, optimization techniques were used to train neural networks to produce behaviour which approximates that of the animal. Although there is no guarantee that the resulting circuits actually exist in those animals, those artificial models pro-vide insight into the issues involved when there is insufficient biological data to construct anatomically realistic models. Where artificial forward-engineered systems have an advantage over strict biological models is in their effectiveness for producing generalizable understand-ing of how behaviour can arise from neuronal systems. Ultimately, we are not par-ticularly interested in how the worm locomotes — we study it because we hope to understand general principles about neuronal systems that we can apply to other bi-ological organisms including humans. There is no guarantee that neuronal circuitry or dynamical mechanisms found in the worm will also be found in other organisms, and in this sense an artificial forward engineered system is on the same footing as a real biological system for producing generalizable knowledge. However, detailed biological studies are extremely time consuming and arduous endeavours: for ex-ample, nearly 30 years of research on C. elegans circuitry still has failed to provide an adequate cellular account of a simple behaviour in a simple organism. In con-trast, artificial systems are relatively cheap and quick to construct and analyze. The selection of a model biological system involves an intuition about the tractability of its analysis and the generality of its mechanism. The same intu-ition can be applied to the selection of an artificial system in order to ensure that it gives insight that generalizes to biological systems. This knowledge transfer can be improved by using models based on physiological mechanisms and expressed 62 in physiological units so they can be readily interpreted by biological researchers working on other systems. 3.2.4 The Human Engineered Approach Previous approaches to engineering non-linear dynamical systems used computer optimization techniques such as recurrent backpropagation (Pearlmut-ter, 1989), genetic algorithms (Beer and Gallagher, 1992), or simulated annealing (Morse et al., 1998). A computer optimized approach has the advantage of finding solutions when the solution space is large and there is little human intuition for guid-ance. However, this is also the weakness of a computer-generated approach when the goal is to develop an intuition for how behaviour can arise from network dynam-ics. Unless the dynamics of resulting circuit is mathematically analyzed (as in Beer and Gallagher, 1992), the mechanism remains mysterious and the designer misses out on the intuition that might be gained by understanding its function. Dynamical systems become extremely difficult to analyze as they increase in size (see Beer (1995) for some of the difficulties involved) and as we progress to larger systems, a computer-generated solution may not be amenable to mathematical analysis. In this chapter, a novel approach was taken instead: a human engineer used intuition about the dynamical processes involved to hand-engineer a modular set of component dynamical systems that were assembled to form the final circuit. As far as the authors know, this is the first attempt to engineer a non-trivial continu-ous dynamical system in this way. A human engineered approach forces the de-signer to identify the issues and develop an intuition about the problem. This chap-63 ter suggests that in spite of the complexities of non-linear interactions, a human-engineered approach might not be so difficult as it seems. To facilitate a human-engineered approach, the design of the dynamical sys-tem was done in steps by building a simple component (the switch) and using it to build more complex components (the oscillator and charger) which were assem-bled to form the complete system. To simplify discussions, the term "dymod" — short for dynamical module — was coined to describe a continuous dynamical sys-tem that is used as a component in the design of a larger dynamical system (which itself is a dymod that could be used to build still larger systems). The term was in-tended to be reminiscent of terms describing computer programming modules such as objects, functions and procedures. It is remarkable that a complex, interdependent, non-linear dynamical sys-tem can be designed in parts and recombined in this way. The design process proceeded according to plan from the original sketch to the final assembly of the dymods into a working system in a surprisingly straight-forward manner. It is in-teresting to speculate whether it is merely a fluke that in this case the non-linear interactions did not destroy the functional properties of the separate modules, or whether this approach is generalizable: are dymods a useful general concept for building non-linear dynamical systems? How scalable is this approach? Because there is no convenient unifying superposition principle for non-linear systems as there is for linear systems, there is no guarantee that it is possible in all cases to combine dymods without destroying their function. The important point here is that we may not necessarily need a guarantee for all cases. It may 64 be sufficient to know how a dymod can be used in a finite set of cases and how to correct for any non-linear interactions in those cases. There are at least two industries that operate on a similar principle to pro-duce extremely large and complex non-linear dynamical systems: the electronics industry, and computer software engineering. They have no appreciable theoreti-cal framework or guarantees (though there is considerable work being done in this direction). Instead of a unifying principle, these fields have a vast collection of tech-niques for solving specific problems (e.g. Knuth, 1997; Horowitz and Hill, 1989). Each technique has a great deal of lore associated with it: when it is appropriate, what pitfalls to watch for, modifications for specific cases, and the success of a so-lution depends in a large part upon the intuition and experience of the designer who has accumulated this lore. The overwhelming success of these industries suggests that a pragmatic, informal approach to combining non-linear modules might be a good starting point in the absence of a unifying theory. If a modular approach to constructing non-linear dynamical systems is pos-sible in a general way, it has a number of advantages over optimization techniques which generate an entire solution at once. Optimization algorithms require the spec-ification of an objective function which characterizes the desired behaviour as a single real-valued "score". This can be sufficient for simple behaviours, but it is uncertain whether complex behaviours or interactions between multiple behaviours can always be characterized so simply. A human-engineered design is limited only by the intuition of the designer and has no requirements that its behaviour be char-acterized in this way. 65 A modular approach may be more scalable than an all-at-once technique since different dymods can be designed and implemented independently. A modu-lar solution is also easier to understand since it provides a decomposition in terms of higher-level building blocks (e.g. compare the dymod view of the reversal cir-cuit in Figure 3.14 with the "flat" cellular view of Figure 3.12). Finally, forward-engineered dymods might suggest higher-level building blocks to use in understand-ing biological circuits. Related work is being done in the emerging field of hybrid systems (Lygeros, 1996) which also seeks to combine non-linear, continuous dynamical systems in a modular way. However, in a hybrid system, the interactions between distinct con-tinuous systems are discrete automata so the result is a hybrid discrete/continuous system. By contrast, the dymod approach seeks to combine continuous dynamical systems into larger continuous dynamical systems and hybrid systems theory does not provide any framework for doing this. 3.2.5 Conclusions and Future Directions The reversal dynamics circuit designed in this chapter is a dynamical sys-tem that governs two discrete behaviours: a persistent forward locomotion and a temporary reversing locomotion. These two behaviours have different underlying mechanisms and the transitions between them are crisp: there is no blending of the two behaviours. Yet both behaviours are governed by a single continuous dynam-ical system consisting of simple tonic non-spiking cells. These behavioural transi-tions occur as the result of bifurcations in the underlying dynamical system. How 66 general is this phenomenon? Can bifurcations explain other forms of behavioural transitions? My further research lies in exploring this question by expanding upon this human-engineered dymod approach. I am constructing a robotic implementation of an artificial creature that exhibits several interacting behaviours, starting with the obvious choices of locomotion, mechanosensation and chemotaxis (using light instead of chemical sensation, in the same manner as Morse et al., 1998). While these behaviours have been well studied in a variety of organisms, I do not seek to mimic the mechanism of any specific organism, but rather to develop mechanisms that make neuroethological sense within the desktop environment and tracked tank-like locomotion of the robot, and relate them to the mechanisms found in other organisms to extract their salient features. This artificial creature approach borrows heavily from the pioneering work of Brooks (1986a), Cliff (1991), Beer (1997), and others and combines the modu-lar behaviour approach of Brooks's (1986b) subsumption architecture with Beer's (1997) dynamical systems approach. The goal of the project is to engineer each behaviour in a modular way, but so that the complete control system is a single continuous dynamical system of biologically plausible cellular mechanisms. This project will also be a good test case for the validity of the human engineered dymod approach since I intend not to resort to automated trial-and-error methods until my intuition completely fails. It is my firm belief that the accumulation of lore and intuition by individual researchers is a necessary precursor to the development of any theoretical frame-67 work for understanding the cellular dynamics underlying behaviour. In fact, it is my suspicion that the brain may not yield to any unifying theory: the accumulated lore may be all we get. The dymod research strategy proposed here is a pragmatic, generative one based on experimentation to determine the conditions under which dynamical sys-tems can be decomposed into modules. To facilitate this strategy, the next chapter presents a framework for implementing and interconnecting dymods in the form of a digital networking protocol called DSS that solidifies the separation of interface from implementation that characterizes dymods. 68 Chapter 4 The DSS Network Protocol for Dymod Implementation 4.1 Introduction A dynamic module (dymod) is a non-linear continuous dynamical system to-gether with a qualitative description of its inputs, outputs, and functional character-istics. Dymods are an approach to constructing and understanding large non-linear dynamical systems in term of smaller modules. They were originally conceived to describe neuronal dynamics in nervous systems (Roehrig and Rankin, 1998) and for experimentally testing their behavioural characteristics (for example, see Fig-ure 4.1), but they may also apply to other control systems governed by continuous dynamical laws. This chapter presents an experimental implementation framework for dy-mods in the form of a network protocol called Digital Signal Sockets (DSS). Dy-69 charger forward oscillator oscillator ^ ^ ^ ^ ^ ^ ^ ^ World Figure 4.1: A neuronal control system for the reversal maintenance behaviour of the nematode C. elegans (Roehrig and Rankin, 1998). It consists of several indepen-dent dymods that are interconnected with sensors and actuators to form a complete behavioural system. The small arrows depict dymod interconnections representing continuous signals. The squares indicate dymods which are governed by non-linear ODEs and the circles represent sensory and motor transducers mods are governed by systems of ordinary differential equations (ODEs) and are implemented using numerical integration techniques. DSS allows dymods to be interconnected with continuous signals and to operate asynchronously using inde-pendent time steps for their internal computation. It also allows these numerical simulations to incorporate live signals from sensors and actuators. Most impor-tantly, the DSS protocol solidifies the modular aspect of a dymod by providing an explicit definition of a dymod's interface inputs and outputs. This work arose out of a prototype neuronal simulator software developed by the author to allow neuronal simulations to be constructed in a modular fashion (i.e. using dymods). The prototype simulator was used to numerically solve the system in Chapter 2, and it handled the explicit definitions of a dymod's input and output interface as well as the mechanisms to send signals between dymods and manage their interconnections. As that work progressed, it became clear that a cleaner and 70 more general approach could be taken by separating the task of numerical simula-tion from the task of managing the interface and communication of signals between different dymods. It also became clear that there was a natural correspondence be-tween dymod interconnections and computer network connections and that for a modest extra effort, a dymod interconnection implementation framework could be cast into a networking protocol to allow large simulations to be distributed over multiple computers, but also to bypass the network if the simulation was entirely contained on a single computer. 4.2 Design Goals The design goals of DSS were: • to solve the problem of interconnecting independent real-time simulations which are numerically computed solutions to ODE initial value problems, • to interconnect simulations using band-limited continuous signals that repre-sent any one-dimensional physical quantity, • to allow live signals to be incorporated into simulations, • to handle all resampling issues that arise when two interconnected simu-lations use unrelated integration step sizes, and provide predictable error bounds, • to multiplex signals over digital computer networks, 71 • to operate with workstations and networks that lack hard real-time facilities (e.g. typical UNIX workstations and TCP/IP networks used by neuroscience researchers), but also to be efficient and have guaranteed performance when these facilities do exist. • to scale well in both number of connections and signal bandwidth. For neu-ronal systems, dymods typically have internal computation step sizes in the order of microseconds and inter-dymod signals have bandwidths in the 1 Hz -1 kHz range. The DSS implementation framework is intended for these low bandwidth, high accuracy signals, but was also designed to scale to higher bandwidth signals as computer and network hardware permit, allowing for the implementation of arbitrarily large dymod systems. 4.3 Review of Existing Frameworks Several existing frameworks were considered as candidates for dymod im-plementation. Quantitative studies of neuronal dynamics are done using numerical neuronal simulators such as Neuron (Hines, 1993) and GENESIS (Bower and Bee-man, 1995). Although parallel and distributed versions are available (Pittsburgh Supercomputing Center, 1998a; Pittsburgh Supercomputing Center, 1998b), these simulators do not support real-time operation and interaction with live signals1. Real-time neuronal studies have been done using silicon artificial neural network (ANN) chips (Hammerstrom, 1995) and a simple bus protocol exists for intercon-1 although the DSS architecture can be readily incorporated in a GENESIS add-on package to add real-time and distributed support in the same straight-forward manner as the mblib GENESIS add-on package (see Chapter 5). 72 necting chips for modular, distributed operation (Mahowald, 1992; Northmore and Elias, 1997). However, these chips typically use limited or inflexible biological models and are complex and expensive to produce. The DIS protocol (IEEE, 1993; Schug, 1995) for performing distributed interactive simulation over the Internet is designed for real-time operation, but it is event-based and intended for military ex-ercises and not suitable for the graded, continuous signals that realistic neuronal simulations require. There are several protocols for distributed digital audio pro-duction (Yamaha, 1998; Young Chang RDI, 1998; Peak Audio Inc., 1998) that are suitable for real-time interconnections of continuous signals, but they are either im-mature or proprietary, and are not well-suited to low-bandwidth signals or TCP/IP networks. Other real-time, distributed networking protocols exist for industrial au-tomation and embedded systems (CAN, 1998; Profibus, 1998) and consumer and professional electronics (1394TA, 1998). These protocols do not provide the re-quired functionality for a dymod implementation framework, but instead may serve as suitable network transmission layers for the higher-level DSS protocol. Of these, IEEE 1394 "Firewire" (1394TA, 1998; IEEE, 1995) is particularly appealing as a network layer because of its high bandwidth, real-time delivery guarantees, and the promise of cheap and widespread availability in the near future. For the TCP/IP networks, there are emerging standards such as RTSP (Real Networks, Inc, 1998) and multicast IP (Comer, 1995), to better handle real-time multimedia transmission over the Internet, and these are also suitable as DSS network layers. The DSS design is independent of the underlying networking layer, but was specifically targeted at two network transport protocols: TCP/IP which is the 73 foundation of the Internet and the desktop workstations used by neuroscience re-searchers, and IEEE 1394 for high-performance dymod implementations with guar-anteed performance. 4.4 Theory of Operation Dymod interconnections represent a shared quantity between two continu-ous dynamical systems. For example, in the coupled system v\ = F1(v1,v2,t) v2 = F2(vuv2,t), (4.1) the variables and v2 represent the shared quantities that couple the system and t represents time. This system is depicted in Figure 4.2 as two dymods with two interconnections. T[vi] and T[v2] represent the transmitted, reconstructed versions of vi and v2 respectively. Figure 4.2: Dymod interconnections in a coupled two-dimensional system. The DSS dymod implementation framework maintains a notion of simula-tion time that is distinct from the actual passage of time, but requires that compu-tation be scheduled and synchronized to the actual wall-clock time. For example, 74 the discrete values wi(ti), ^1(^2), • • • produced by the JFi dymod must be delivered to the F 2 dymod within some interval of the actual times ti, i 2 , • • • which are taken relative to some starting time. There are several distinct problems to be addressed in a dymod framework: signal and time representation, signal reconstruction, connec-tion management, time synchronization, real-time scheduling, and network trans-mission. The DSS protocol deals with signal and time representation, signal re-construction and connection management. Solutions to the other problems are also discussed in the following sections. 4.4.1 Signal and Time Representation DSS version 1 is experimental and deals with only one type of signal: a continuous real-valued signal that is bandlimited (containing no frequency compo-nents above a highest frequency). The Nyquist sampling theorem (Oppenheim and Schafer, 1989; Nyquist, 1928), states that bandlimited signals can be exactly rep-resented by a discrete series of samples taken at a sampling rate of at least twice the highest frequency. This minimum sampling rate is called the signal's Nyquist frequency. DSS version 1 represents these signals by a discrete set of 32 bit values, but in future versions, this can be expanded to include other signal encodings. DSS uses an absolute time reference for signal data, and each signal sam-ple is timestamped with the absolute time at which that sample was taken. Using a timestamp allows a receiving dymod to recover the time the sample was taken without needing to account for the latencies and variabilities of network transmis-sion. By using an absolute time reference, the DSS protocol is simplified because 75 the time synchronization problem can be solved by well-known time synchroniza-tion protocols (Mills, 1992). This lightens the requirements for the DSS network transmission mechanism so that DSS messages can easily be bridged between dif-ferent network types (such as TCP/IP networks and the higher-speed IEEE 1394 isochronous network that has its own intrinsic clock) without requiring network-level clock synchronization. In DSS version 1, the timestamp is a 64 bit fixed-point quantity representing the time in seconds since the epoch of midnight January 1, 1970 Coordinated Uni-versal Time (UTC) and is computed according to the POSIX.l get t imeofday system call (IEEE, 1996). Note that the value returned by this call is computed from the current time of day and ignores leap seconds. While this is adequate for experimental purposes, is discontinuous whenever a leap second is inserted into UTC (roughly every two years) and future DSS versions may incorporate a differ-ent encoding. For a more efficient and compact representation, each DSS connection main-tains a 32 bit reference time in whole seconds — the connection epoch — as the zero point for transmitted signal timestamps. The absolute timestamp is obtained by adding the sample's transmitted timestamp to this epoch. The transmitted times-tamps are also represented in 32 bits and can only represent a fixed interval of time: in DSS version 1, the timestamp represents microseconds since the connec-tion epoch, and will overflow the 32-bit quantity approximately 72 minutes after connection establishment. Before this occurs, the connection epoch is renegoti-ated as described in Section A. 1.2. Future DSS versions may incorporate higher 76 resolution timestamps. 4.4.2 Signal Reconstruction Dymod implementations operate independently and asynchronously, typi-cally with internal computation rates that are higher than the signal interconnection sampling rates. Incoming signal samples must therefore be used to reconstruct the continuous signal in order to avoid step discontinuities between sample values. This reconstructed continuous signal is then resampled at the appropriate intervals for computation in the receiving dymod. In Figure 4.2, the transmitted versions of a signal v is given by T[v]. The transmission functional T involves two undesirable components: delay and approx-imation. It can be written as: T[v](t) = v(t-5) + e(t) (4.2) where 8 is the delay incurred by signal transmission and reconstruction, and e(t) is the error in the reconstruction. There is a tradeoff between the amount of delay and the amount of inter-polation or prediction error. For instance, in the system (4.1) it is possible to use a zero delay by fitting a polynomial to the samples of v\ for t < tn and use it to predict values for v\ in the interval [£„,t„+i]. This is a similar approach to the predictor-corrector numerical method for ODE integration (Gear, 1971), except for the corrector step: because the solution to (4.1) is distributed, the correction of v\ via the computation of F\ cannot be applied immediately to the computa-tion of v2(tn+i), but must wait instead until the computation of v2{tn+2) when a 77 new v\ sample has been received from the F\ dymod. For smooth functions, the predictor-corrector method is highly accurate and this modification may still yield good results. However, the formal error analysis would be complex and this modi-fied predictor-corrector approach is left for a future endeavour. In some cases a delay is acceptable. If the physical system includes delays, a model can incorporate them so that the dymod interconnection delay plays a part in the model. This is the case for neuronal systems when synaptic transmission and axonal spike propagation involve delays. If a delay is used, no prediction is required and the problem is a simpler one of interpolating an intermediate signal value from data samples on either side. For bandlimited signals, there is a well-established theory and a wealth of efficient numerical techniques for doing this (Crochiere and Rabiner, 1983). A bandlimited signal can be reconstructed exactly from its sequence of samples using the well-known sine interpolation function (Oppenheim and Schafer, 1989) oo *>(*) = E vnsmc(n(t-tn)Fs), (4.3) n=—oo where Fs is the signal's Nyquist frequency, tn is the time value for sample n, vn = v(tn) are the sample values, and sinc(a;) = sin(x)/x. In practice, there are different techniques for truncating this doubly-infinite sum and implementing it using efficient digital reconstruction filters (Crochiere and Rabiner, 1983). For ex-ample, a A;-tap finite impulse response low-pass digital filter uses a symmetrically truncated sum of k terms and incurs a delay of ^ in the signal pathway. If an incoming signal has a higher bandwidth than the internal computation rate of the receiving dymod, the signal should be filtered to remove the higher fre-78 quency components which would cause aliasing — a form of noise caused when higher frequencies obscure the low frequency trends when sampling the signal. In this case, Fs in (4.3) should be set to the Nyquist frequency of the receiver's band-width capabilities. The reconstruction accuracy is determined by the design and width of the reconstruction filter. Better accuracy is obtained by using more input samples with a wider interpolation filter kernel, but at the expense of a longer delay. Figure 4.3 shows how reconstruction accuracy changes with filter width and delay. 0.001 1 ' ' ' ' 1 1 ' ' 1 1 2 3 4 5 6 7 8 9 10 Zero Crossings Figure 4.3: Reconstruction Error vs Filter Width. This shows typical peak inter-polation errors in upsampling a spectrally-white test signal to 14.3 times its orig-inal sampling frequency Fs (n=1000 interpolations). The upsample ratio of 14.3 was chosen to avoid a simple rationally related sampling rate change, but these interpolation errors are typical of other ratios. Filter width is given as the num-ber of zero crossings on either side of the sine reconstruction kernel included in a Ffann-windowed filter design. The test signal was bandlimited to Fs/4 so that it was in effect oversampled by a factor of 2. The filter delay can be computed by delay = zero crossings/Fs, and the approximate number of samples used by # samples — 2 x zero crossings + 1. The average errors were less than 0.00002% when at least 3 zero crossings were used. In this experimental DSS version, a simple Hann-windowed filter design 79 (Oppenheim and Schafer, 1989) that included three zero crossings on either side of the sine interpolation kernel was used, with a resulting delay of three sample periods. The reconstruction filter was computed using double-precision floating point arithmetic and was not optimized for minimum error or performance. Win-dowed filter designs such as this one yield good frequency-domain characteristics and efficient implementations, but do not produce optimal time-domain interpo-lations. Filter design techniques to minimize the maximum or least-squares error in the time domain are provided in Crochiere and Rabiner (1983). For a more efficient fixed-point arithmetic interpolation filter that supports non-uniform resam-pling and is suitable for hardware or embedded DSP applications, see Smith and Gossett (1984). Reconstruction accuracy is also affected by "lost samples" which occur when a sample has not been received in time to be used in the signal reconstruction. This can happen for three reasons: real-time scheduling problems, time synchro-nization errors, and network delay. 4.4.3 Real-time Scheduling Accurate signal reconstruction is critically dependent upon accurate real-time scheduling of computation. If a receiver's computation begins too early, the samples required for computation may not have arrived; if it begins too late, the computation might not complete in time resulting in a delay (and more lost samples downstream). These lost samples result in a reduction in the signal reconstruction accuracy and possibly exceeding the error tolerance of the application. This can be 80 mitigated by oversampling the signal (sampling it at a rate greater than its Nyquist frequency) so that even if occasional samples are lost, the Nyquist criterion is still satisfied. A dymod's computation of a value v(U) at time tt must be scheduled to commence at time in order to incorporate any live signal values at time U. In order to keep up with real time, this computation must be completed by time ti+i = ti+dt where dt is the dymod's internal step size. If 8c is the time taken for internal computation, then Sc/dt is the dymod's "real-time load". When v(t) is transmitted to another dymod, its transmitted version is T[v] (t) = v(t—5)+e(t). In order to ensure that no samples are lost, the receiver must schedule the computation that makes use of T[t>](i) to occur in the interval [t, t + Ts] where Ts is the sampling interval of the transmitted signal v and, in addition, the signal delay 8 must satisfy 8> 8R + 8C + 8N, (4.4) where 8R is the reconstruction delay imposed by the reconstruction filter, 8c is the computation time taken to produce samples of v(t), and 8N is the network trans-mission time. Note that 8R is determined by the reconstruction filter design char-acteristics while 5c and 8^ are empirical quantities determined by the speed of the computer and network hardware. Faster hardware can reduce 5c and 5N, but only a design change in the reconstruction filter can reduce 8R. In our experimental DSS version 1, the filter delay 8R is negotiated at con-nection time based on the sender's and receiver's sample rate and filter width, and the signal's delay 6 is set to 5R with no provision for 8c and 8N. This does not 81 present a problem providing 8c + 8N < Ts, in which case at most one sample will be lost per reconstruction. In our experiments, we compensated for this by over-sampling by 2 times. The signal delay is made available to the application via the DSS API (dss_get inf o, Section A.2) Future versions of DSS will provide the means to establish and monitor 8c and 8N in addition to 8R. Embedded systems typically have excellent real-time facilities, but UNIX-based dymod implementations are more problematic. The Posix s e t i t i m e r sys-tem call is typically used to schedule real-time activity on UNIX systems, but real-time guarantees vary widely between UNIX implementations. Sun's Solaris and SGI's IRIX are considered real-time operating systems (Real-Time Magazine, 1998) with guarantees on real-time performance, but that is somewhat moot con-sidering the lack of any real-time guarantees for TCP/IP networks. A UNIX and TCP/IP dymod implementation is intended for experimenting with the qualitative behaviour of systems and should not be relied upon for guaranteed numerical ac-curacy. However, the performance of UNIX systems and TCP/IP networks can be quite adequate for behavioural experiments providing that the system limita-tions are tested and determined in advance. Some UNIX implementations — on PC hardware in particular — have poor real-time clock support with resolutions in the tens of milliseconds, and any more precise scheduling is impossible. An em-bedded DSP (digital signal processing) implementation using a real-time network such as IEEE 1394 can be used for guaranteed numerical accuracy. For a good practical starting point on real-time computation issues, see the online Real-Time Encyclopaedia (Real-Time Magazine, 1998). 82 4.4.4 Time Synchronization DSS requires accurate time synchronization between sender and receiver, but does not provide a synchronization mechanism and instead relies on an exter-nal protocol. On TCP/IP-based UNIX implementations, the NTP protocol (Mills, 1992) is used to synchronize the workstation clock to a reference signal obtained by radio or the Internet. In embedded implementations using isochronous networks, a hardware clock synchronization mechanism can be used. In order for an embedded implementation to interact with a TCP/IP implementation, the embedded network clock master should be synchronized to Universal Coordinated Time. The accuracy of time synchronization limits the real-time scheduling accu-racy and therefore the signal reconstruction accuracy as described in the previous section. Most typical UNIX installations can maintain a synchronization accuracy to within a few tens of milliseconds using NTP (Mills, 1990). This time synchro-nization error is the major factor that limits signal bandwidth when using DSS in a UNIX and TCP/IP environment. Sub-millisecond accuracy is possible using special kernel patches and extra hardware (Mills, 1994). In our experimental test network, system clocks were synchronized with an accuracy of approximately 10 milliseconds and we experienced nominal recon-struction errors when the sampling interval was substantially greater than the time synchronization error, but the reconstruction errors increased dramatically to 100% as the sampling interval approached the time synchronization error and the result-ing real-time scheduling errors caused samples to be lost from the reconstruction (Figure 4.4). 83 Peak Error (%) 5 10 15 20 25 30 36 40 45 60 Sampling Interval (ms) Figure 4.4: DSS Reconstruction Errors vs Sampling Interval. This shows how the peak interpolation error changes as the sampling interval approaches the time synchronization error (approx 10 ms) and the resulting real-time scheduling errors cause samples to be lost from the reconstruction. A spectrally-white test signal that was bandlimited to Fs/4 was used and it was upsampled at the receiver by a factor of 14.3. Three zero crossings were used for the reconstruction filter resulting in nominal reconstruction errors of 0.2% (cf. Figure 4.3) when no samples were lost. 84 4.4.5 Connection Management DSS signal connection endpoints are called DSS ports and they are identified by a DSS address (see Section A. 1.1). DSS connections carry unidirectional sig-nals and therefore ports must be of either input or output type. This directionality only refers to signal data transmission; all ports can send and receive DSS con-trol messages which are used to establish and remove connections between DSS ports. Both ports must exist and be active before the connection can be made. The use of DSS control messages for connection management allows dymods to be im-plemented independently of any controlling user interface: for instance, a dymod implementation can be a featureless black box with a network jack, and a DSS-enabled web browser could be used to control its behaviour and to connect it to other dymods. To facilitate connection management, the DSS protocol includes a name service to associate symbolic names with DSS ports. This allows a user inter-face to specify connection endpoints by name rather than their DSS addresses. In future versions, the name service may also provide additional information such as the physical quantity and units the signal represents. 4.4.6 Network Transmission The DSS protocol does not include a network transmission mechanism and is meant to be layered on top of an additional networking protocol (see Figure A. 1). DSS requires two kinds of network transport services: an unacknowledged iso-chronous datagram service for delivering sample data at regular rates, and an asyn-chronous acknowledged datagram service for control messages. This experimental 85 DSS version was implemented on top of TCP/IP using UDP (Comer, 1995) for both isochronous and asynchronous services. Future versions may be implemented us-ing the emerging real-time streaming protocol (RTSP) (Real Networks, Inc, 1998) for TCP/IP. Typical 10 Mbit Ethernet LANs can saturate at 4 Mbit/s data rates. The DSS isochronous signal messages are 20 bytes (Section A. 1.2), the IP and UDP headers are 28 bytes and the ethernet frame overhead is 26 bytes (Comer, 1995), yielding a maximum total signal bandwidth of a few kHz, which is sufficient for simple neuronal simulations. IEEE 1394 provides isochronous and asynchronous services intrinsically and is designed with scalable data transfer rates with current rates of 100, 200 and 400 Mbit/s. With a more compact 12 byte isochronous frame overhead (IEEE, 1995), IEEE 1394 can theoretically support a total DSS signal bandwidth of over 1 MHz. The DSS protocol specification and application programmer interface (API) is presented in Appendix A. 4.5 Conclusions The experimental DSS implementation described here suggests the feasibil-ity of using digital networks to interconnect continuous dynamical systems modules (dymods) for distributed real-time simulations. DSS solidifies the dymods' separa-tion of interface from implementation by providing an explicit mechanism to define the interface inputs and outputs via DSS input and output ports. DSS provides a 86 standard mechanism to interconnect dymods ports with band-limited continuous signals, and also provides a name service to allow a remote connection manager to identify dymod ports by name. The DSS protocol is targeted at two network architectures: TCP/IP for low-cost, low-bandwidth simulations and IEEE 1394 for high-bandwidth simulations with guaranteed performance. With UNIX and TCP/IP implementations, the total signal bandwidth is theoretically limited to a few kHz on typical 10 Mbit ethernet networks, but in practice this is limited by the accuracy of time synchronization across the network. In order to use signals with bandwidths of greater than about 50 Hz, special UNIX kernel modifications and hardware are required for accurate time keeping and network time synchronization. The current experimental implementation requires a hard delay along dymod interconnections which limits its utility for solving general systems of ODEs such as the C. elegans dymod example (Roehrig and Rankin, 1998). However, an extrap-olation algorithm may be used to implement a modified predictor-corrector numer-ical method for solving general ODE initial value problems without delays. The next step in the development of the DSS dymod implementation framework is to perform the error analysis and implementation for this modified predictor-corrector method. DSS provides only a basic implementation framework on which to build other dymod tools such as neuronal simulators, and robots for experiments in com-putational neuroethology using dymods. The next chapter presents an inexpensive robot system to be used for dymod experimentation. 87 Chapter 5 A Desktop Robot System for Experimental Neuroethology using Dymods 5.1 Introduction The dymod approach of Chapter 3 is intended for constructing complex neu-ronal models of behaviours and to ultimately allow neuroscientists to create com-plex behavioural models using high-level building blocks that are grounded in a physiological implementation, but without having to understand all the physiolog-ical details. The DSS network protocol of Chapter 4 provides the means to inter-connect dymod implementations using typical UNIX workstations and TCP/IP net-works, scales to support arbitrarily large dymod systems, and provides an explicit mechanism to define a dymod's interface. 88 This chapter adds to the dymod tools by presenting a robot system that the author has constructed in order conduct neuroethology experiments. It continues the theme of the previous chapters with the goal of empowering the general neuro-science researcher, in this case to perform robot neuroethology experiments without requiring expertise in robotic engineering or real-time numerical computation. 5.2 Design Goals The robot was designed to meet the following goals. To be accessible to gen-eral neuroscience researchers, the robot needs to be inexpensive and easy to con-struct. It should not require any special expertise in electronics or robotics design. It should be easy to use so that neuroscientists can quickly experiment with ideas with a minimal of complication by irrelevant technical issues. The robot should be flex-ible and adaptable to different neuroethological tasks in order to appeal to a wide neuroscience audience. For instance, a chemotaxis simulation could use simple in-frared or optical sensors mounted to the robot and use lights of different colours to represent different chemicals. Finally, to allow easy comparison between different levels of behaviour models, the robot control system should allow for testing of both simple neuronal models and quite complex and realistic models. 89 5.3 Chassis Design 5.3.1 Locomotion A wheeled design was chosen because it is simple, robust, and requires few motors and mechanical components. A disadvantage of a traditional wheeled car chassis is that the vehicle must perform a 3-point turn to reorient it self in place. This is a fairly complex task that doesn't have much neuroethological relevance and would require a substantial neuronal control system. To avoid this problem, two approaches were considered. The first was a tricycle design with a tail-dragger wheel (see Figure 5.1). Two motors drove the main left and right wheels, and a third free-rotating wheel at the tail provided support and proprioceptive feedback. The two drive mo-tors operated independently to move the vehicle or turn it left or right. The rear wheel was mounted on a swing arm which pivoted as the robot turned. The swing arm was mounted on the shaft of a 360° precision servo potentiometer to provide proprioceptive feedback on the tail's angular orientation. The rotational speed of the rear wheel was to provide proprioceptive feedback as to the vehicle's motion, but it was difficult to construct a sensor on the small tail-wheel swing arm as well as rig the wiring across the freely-rotating shaft it was mounted on. This design was unnecessarily complex and was abandoned in favour of a tank-like design (Figure 5.2) which suggested itself by the presence of tank treads in a LEGO Dacta kit. The tank chassis is much simpler, requiring only the two drive motors and is able to orient itself in place by driving its two treads in opposite 90 r\IK)l&OnfU> MOXfOT>WGi | f0U=S . Figure 5.1: Preliminary Tail-dragger Robot Design 91 Figure 5.2: LEGO Tank Robot Design directions. 5.3.2 Proprioception and Sensors Proprioception is an important part of a control system. Most biological motor control operations require sensory feedback from stretch receptors to operate properly. Robotic limbs also use sensors to provide information on the joint angles and stresses. In a wheeled vehicle, proprioception does not have a direct analogue in the biological world, but is nevertheless important. If the robot is driving a motor in order to locomote, it is important to know whether the motor is actually turning. The robot may have run against an obstacle, or encountered more difficult terrain requiring more motor power. More importantly, if power is being delivered to a motor but the motor is not turning, it draws considerably more power than if it were turning and can uselessly drain valuable battery energy. 92 There are two possible signals that can be used as proprioceptive feedback from a wheeled vehicle: the angular position of each wheel, and the rotational speed of each wheel. The latter was chosen as being generally more ethologically relevant since the wheel's absolute position is typically unimportant. However, there may be some tasks involving slow or precise positioning where absolute wheel position is more important. Several ways of encoding wheel speed were explored. A small DC motor from a slot-car racing car was geared to the wheel. As the wheel turned, the DC motor generated a voltage proportional to the rotation speed. However, the motor needed to be geared up from the wheel in order to provide a sufficient voltage and the geared-up inductive resistance of the motor was sufficient to prevent the wheel from rotating as the vehicle moved. A larger, heavier vehicle might be able to overcome this frictional resistance, but it would not be as suitable for a desktop robot. Optical shaft encoders use a photocell to detect a light beam passed through a wheel containing slits. By measuring the time difference between on and off light pulses, the rotational speed of the wheel can be determined. The existing design does not use an optical shaft encoder, but rather a resistive one (manufactured by ALPS) that was cannibalized from an old Microsoft mouse. However, it has proven to be noisy and somewhat unreliable and is now difficult to obtain. The next revision will use Bourne optical sensors. Additional sensors can be used depending on the neuroethological experi-ments to be done. 93 5.3.3 Construction The chassis was constructed from LEGO Dacta (Technics) using kits 9605, 8826, as well as some other assorted LEGO parts. The chassis has a compact gearing mechanism to gear the motors down by a ratio of 25:1 to provide adequate torque to the treads. The design also houses both the microcontroller and a battery compartment below the tank's deck, providing a clean base for additional sensory and actuator mechanisms for various neuroetho-logical experiments. The battery compartment holds 4 C-cell batteries which can deliver a typical maximum load of 500 mA (with 2 motors fully activated) for 14 hours (using alkaline batteries). The construction diagrams are given in Appendix B. 5.4 Controller To allow for arbitrarily complex and realistic neuronal control systems, the controller was designed so that the detailed neuronal simulations would be com-puted by outboard high-speed computers that are connected to the robot with a wire tether, while an inexpensive onboard microcontroller handles the simpler task of managing the sensory data acquisition, motor control, and host communication. This approach was taken because there are two different tasks that need to accomplished: detailed, timing-sensitive control of sensors and motors, and high-speed numerical computation. These tasks have different computational require-ments and there are inexpensive generic solutions to each problem, but a single 94 solution to both would require special-purpose hardware. The limitation of having to tether the robot to a computer was not seen as a significant one for a desktop system. While the recent availability of inexpensive high-performance 32-bit micro-controllers makes it feasible to build stand-alone autonomous robots which perform all neuronal computations onboard, a tethered solution has the advantage of being able to use a workstation to monitor, analyze and modify the neuronal operating parameters during behaviour. 5.4.1 Overview Fred Martin's MIT Miniboard (version 2.1) was chosen for providing on-board sensor and motor control. (Martin, 1995) This is a compact, low-power, in-expensive ($50) board based on the E2 variant of the popular Motorola 68HC11 microcontroller chip, and fits nicely into the chassis of the LEGO robot. The 68HC11E2 has 2 Kbytes of on-chip read-only program memory and 256 bytes of R A M which is sufficient for the control and communication program (but would be wholly inadequate for numerical computation). A communication protocol was designed to allow communication of sensor and motor signals between the Miniboard and the host computer over a serial cable. The protocol was implemented in the control program of the Miniboard as well as in a portable C library for use on the host computer. This allows any standard C program to interact with the robot, and gives maximum flexibility in implementing computational neuronal models for use with the robot. An add-on module for the 95 popular GENESIS neuronal simulator was created to allow the use of GENESIS simulation for real-time control of the robot. The Miniboard provides the robot with the capability for eight analog (grad-ed) sensors, eight digital (on/off) inputs or outputs, four higher-power motor out-puts, and two timer inputs for response to time-sensitive events. 5.4.2 BINMON: The Miniboard Control Program The miniboard control program (called BINMON) is responsible for several tasks. • It reads and records the robot's analog and digital sensor values every mil-lisecond. • It maintains an accurate millisecond clock to use as timestamps. • It implements the host communication protocol to communicate sensor and motor signals. • It modulates the motor output signals to provide different motor speeds using pulse-width modulation (PWM). • It decodes the timer input signals to implement a simple shaft-encoder sup-port. The control program was based on Fred Martin's original HEXMON code, but was substantially rewritten to support the communication protocol and resistive shaft-encoder support. Full details are included in the m b l i b package (see below). 96 Communication Protocol The communication protocol provides a mechanism for the host computer to communicate with the Miniboard. This protocol operates in two modes: an "ASCII" (human-readable) mode which can be used for diagnostics and a binary mode for efficient communication of sensor and motor signals. It is important for the sensor data to be accurately timestamped in order to maximize numerical accuracy when simulating the neuronal controller model. Rather than try to synchronize the Miniboard's clock with the host computer's clock, a simpler approach was taken. The Miniboard operates in a passive mode, recording sensor data every millisecond, but not initiating any communication. When it receives a request from the host computer, it responds by transmitting the sensor data. The host computer can timestamp the sensor data when it receives it, and can correct for the transmission delay to accurately determine the sensor acquisition time. This mechanism fits well with existing ASCII command-response mode of the original HEXMON code. The original ASCII commands were preserved mostly unchanged for debugging purposes. The ASCII commands return a human-readable response which is terminated by a command-prompt' >'. The ASCII commands are summarized as follows: s Perform a reset and resynchronize to the host. rmmmm Read byte at location mmmm. wmrnmmdd Write byte dd at location mmmm. 97 qmmmm Read word at location mmmm. zmmmmdddd Write word dddd at location mmmm. v Print monitor version. d ASCII dump of state. The binary communication protocol is initiated by the host by a b command. This causes the Miniboard to receive a frame of control commands and once it has been received, to return a frame of sensory data. Full-duplex operation (i.e. trans-mitting the sensory data simultaneously with receiving the control commands) is not possible due to the specifics of the Miniboard's serial communication hardware implementation (see (Martin, 1995)). The control command frame consists of 11 bytes: a motor control byte (4 bits on/off, 4 bits direction), 4 motor speed words (each consisting of a 16-bit PWM mask), a data direction control byte for the digital input/output port, and a data byte for that port. The sensory data frame consists of 11 bytes: one byte for each of the eight analog inputs, a data byte for the digital in-put/output port, and a byte for each of the two shaft encoder inputs. The transaction is terminated by the command prompt character. Timing Discussion Since computer operations occur many orders of magnitude faster than the communication time over a serial cable, it is necessary to correct for serial transmis-sion time to obtain an accurate timestamp for the sensory data. The binary trans-mission time can be computed by the number of bits of information transmitted 98 divided by the bit rate. At the current bit rate of 9600 baud, command transmission (12 bytes of 8 bits with 1 stop bit) takes 11.25 milliseconds. Because of peculiarities of the Miniboard's serial communication hardware implementation, all transmitted bytes are echoed back to the host. In addition to these bytes, the return frame con-sists of 11 bytes for a total of 23 bytes that must be received by the host (there is no reason to wait for the command prompt). This reception takes 21.56 milliseconds (but because of the hardware echo, this happens concurrently with transmission of control information). Therefore, the Miniboard will respond with motor commands with a 11 ms delay from the issuing of the motor commands, and the sensory data from the Mini-board will only be able to be used by the numerical simulation 22 ms after it is actually acquired. (Because the Miniboard only updates at a rate of 1 kHz, these delays are only accurate to within 1 ms.) DSS can be used for the numerical simulation by embedding the Miniboard sensor values into DSS packets. If this is the case, the DSS timestamps can be cor-rected for this delay and the Miniboard sensory signals can be processed accurately. However, as in the case of DSS network transmission latencies, this delay imposes a bandwidth restriction on the signals which can only be overcome by faster trans-mission time. (A 22 ms delay means a maximum signal bandwidth of about 20 Hz.) See below for a brief discussion on how to increase the Miniboard's serial transmission speed. 99 Miniboard Host C Library A library of routines was implemented in the C programming language to allow a host computer to interact with the Miniboard using the communication pro-tocol. The C library maintains a copy of the miniboard's state on the host computer and periodically synchronizes this state with the Miniboard. The library application programmer interface (API) is given in Appendix B. The m b l i b Package The m b l i b package contains implementations of the BINMON Miniboard control program and Miniboard C Library, together with the following additional components: mbview An interactive miniboard viewer to allow viewing and manipulation of the Miniboard. It has a simple keyboard interface and a text-based display to allow it display on any VT100 compatible terminal. It also displays various timing measurements to analyze serial port latencies. d l m l l A UNIX port of the Miniboard program downloader. GENESIS Miniboard Library This is a library module for the popular GENESIS neuronal simulator to allow it to synchronize the simulation to real-time and interact with live signals from the Miniboard. The m b l i b package can be obtained from http://www.crispart.com/mblib. 100 5.5 Future Work There are many avenues for continued work on this robot project. At the forefront, is the implementation of a Miniboard DSS interface to encapsulate the Miniboard signals into DSS packets. This will allow the Miniboard to interoperate with other DSS modules. The implementation would be in the form of a daemon process that runs a continuously updating Miniboard on a serial port and maintains a collection of DSS ports for each Miniboard signal. In addition, a GENESIS DSS module needs to be implemented which would allow the GENESIS simulator to interoperate with any DSS signal. The BINMON Miniboard Control Program needs to be modified to support optical shaft encoders. The Bourne Model ENC1J-D28-L00T28 shaft encoder is a compact low-friction shaft encoder that can be used with appropriate modifications to the BINMON program. Modifications to the LEGO robot to accommodate this encoder also need to be done. In addition, the binary communication protocol can be made more compact. Currently, the motor PWM masks are 16-bits which means 8 bytes are transmitted every frame. However, these PWM masks only encode 16 different speeds according to a table kept on the host. This table could be moved to the Miniboard and the 4-bit motor speeds could be transmitted instead for a savings of 6 bytes (6 ms at 9600 baud). To substantially improve the signal bandwidth between the Miniboard and the host, the serial speed needs to be increased. The maximum standard speed of the Miniboard is 9600 baud using the standard 8 MHz clock, but this clock could be replaced by a 4.9152 MHz crystal to obtain higher standard baud rates (such as 101 38.4 Kbaud) at the expense of reducing the Miniboard CPU's bus frequency from 2 MHz to 1.2 MHz. However, the Miniboard is not performing any computationally intensive work, and this loss of CPU speed will likely not be a problem. At 38.4 Kbaud with the savings from the PWM modifications, the host-Miniboard update time would be reduced from its current 22 ms to 4 ms, with signal bandwidth of 125 Hz. 5.6 Conclusions This chapter presents a desktop robot system the author has constructed for conducting neuroethology experiments using dymods. The inexpensive LEGO robot uses a wheeled-design for simplicity and reliability. It uses a tank-like chassis with treads, which gives it the ethologically relevant ability to orient in place, un-like other car-like designs which require three-point turns. The chassis is compact and houses the battery compartment, motors and computer control system below the tank's "deck" to provide maximum flexibility for adding sensory and actuator apparatus. The control system is a tethered design: an MIT Miniboard monitors and controls the robots sensors and motors and transmits them along a cable to a host UNIX computer which performs the actual neuronal computation. This chapter also presents a Miniboard program (BINMON) that performs the control functions, a UNIX library to communicate with the robot via serial cable, and a library add-on for the popular GENESIS neuronal simulator to allow it to communicate with the robot. Currently, the LEGO robot is fully mobile and contains proprioceptive sen-102 sors for tread motion, and the microcontroller's other 6 sensory inputs and 2 motor outputs are not utilized and available for expansion. The support software has been developed to the point where a simple single-compartment neuron simulation run-ning under GENESIS was able to successfully control locomotion in the robot. 103 Chapter 6 Conclusions This dissertation introduced the dymod concept to help account for a pre-viously unexplained behaviour in the nematode C. elegans: its ability to continue swimming backwards for a period of time after the end of a tap stimulus. The dy-mod approach provides building-blocks for constructing a dynamic behaviour out of a set of simpler dynamical structures. Current brain theories (Kelso, 1995) speculate that understanding dynamical structures is the key to understanding how the brain generates abstract behaviours like language, planning and perception. However, so far we have no building blocks to quantitatively describe dynamical structures except at the most detailed level of the physiological components: cells, channels, etc. It is unlikely that we will be able to understand as complex a behaviour as perception in terms of cells and channels without some form of higher-level abstract building blocks. The dymod approach is a step towards those building blocks. The purpose of the dymod framework is to describe a dynamical system in 104 terms of a set of simpler dynamical modules. This dissertation illustrates the dymod concept with a simple example and lays out a research programme to investigate the generality and usefulness of the dymod approach. The research programme is based on the generative computational neuroethology approach pioneered by Beer, Cliff and others (Beer, 1997; Harvey, Husbands, Cliff, Thompson and Jakobi, 1997) using robots to ground neuronal models in a complete behavioural system that in-cludes its environment as part of the system. The dissertation also presents a pre-liminary set of tools for embarking upon this research programme: the DSS dymod implementation framework, and a general-purpose desktop robot platform. 6.1 Future Directions The next step in the research programme that was laid out in Chapter 3 is to combine the locomotory circuit of Chapter 3, the implementation framework of Chapter 4 with the robot of Chapter 5 (see Figure 6.1) To complete the next step, the following projects need to be completed. • The DSS protocol must implement the modified predictor-corrector numer-ical method (Section 4.4.2) to allow signal reconstruction without hard de-lays. The tonic and graded synaptic model for C. elegans (Section 2.6 and Section 3.1.1) does not use hard delays — indeed many neuronal dynami-cal systems do not use hard delays — and the current DSS implementation cannot be used for these models which substantially limits its utility. The implementation of the modified predictor-corrector numerical method would require a fair amount of detailed numerical analysis to determine convergence 105 DSS CONNECTION Figure 6.1: The next step in the research programme is to combine the locomo-tory circuit with the DSS implementation framework and the robot. A robot with a tap sensor is to be tethered to a computer running a DSS-enabled m b l i b (Sec-tion 5.4.2) to implement a dymod representing the robot's sensors, motors and its environment. It will be connected to a second computer implementing the oscilla-tor dymod which would in turn be connected to a third computer implementing the charger dymod. The dymods can be implemented using a DSS-enabled version of the GENESIS neuronal simulator (Section 5.4.2). 106 properties and error bounds. • DSS-enabled implementations of mbl ib and GENESIS need to be com-pleted. Both are relatively straight-forward modifications of the existing im-plementation presented in Section 5.4.2. .2 Novel Contributions The novel contributions of this dissertation are summarized as follows: 1. A physiologically detailed cellular model of the nematode tap withdrawal circuit (Chapter 2). 2. A graded and tonic synaptic model and a treatment of the distinction between a cell's resting potential and its in-circuit steady-state potential (Section 2.7). 3. A cellular account of how the nematode might continue its reversing after the end of a stimulus using neuronal network dynamics (Chapter 3). 4. Dymods: a way of decomposing dynamical neuronal structures into modular subcomponents (Chapter 3). 5. DSS: an implementation framework for dymods that is designed to scale to handle large, complex systems of dymods (Chapter 4). 6. A hybrid numerical method for solving ODE initial value problems in real-time using digital signal reconstruction filters as part of the numerical method (Section 4.4.2). 107 7. An inexpensive, general-purpose desktop robot system intended to allow neu-roscience researchers to experiment with the dynamical basis of behaviour without requiring expertise in robotics or numerical solutions to systems of ODEs (Chapter 5). 8. A research programme for investigating the modular properties of dynamical neuronal structures and how they interact, together with a preliminary set of tools for embarking upon that programme. 108 Bibliography 1394TA (1998). IEEE 1394 trade association. URL: (http://www.1394ta.org) [accessed Apr 7, 1998]. Abbott, L., Marder, E. and Hooper, S. (1991). Oscillating networks: Control of burst duration by electrically coupled neurons, Neural Computation 3(487-497). Achacoso, T. B. and Yamamoto, W. S. (1992). AY's Neuroanatomy ofC. elegans for Computation, CRC Press, Boca Raton, FL. Avery, L., Raizen, D. and Lockery, S. (1995). Electrophysiological methods, in H. Epstein and D. Shakes (eds), Caenorhabditis elegans: Modern Biological Analysis of an Organism, Academic Press, San Diego, pp. 251-269. Beer, R. D. (1995). On the dynamics of small continuous-time recurrent neural networks, Journal of Adaptive Behavior 3(4): 469-509. Beer, R. D. (1997). The dynamics of adaptive behavior: A research program, Robotics and Autonomous Systems 20: 257-289. Beer, R. D. and Quel, H. J. (1993). Simulations of cockroach locomotion and escape, in R. D. Beer, R. E. Ritzmann and T. McKenna (eds), Biological Neu-ral Networks in Invertebrate Neuroethology and Robotics, Academic Press, pp. 267-285. Beer, R. D. and Gallagher, J. C. (1992). Evolving dynamical neural networks for adaptive behavior, Adaptive Behavior 1: 91-122. Bennett, M . V. L. (1972). A comparison of electrically and chemically mediated transmission, in G. D. Pappas and D. P. Purpura (eds), Structure and Function of Synapses, Raven Press, New York, pp. 221-256. 109 Bialek, W., Rieke, E, de Ruyter van Steveninck, R. and Warland, D. (1991). Read-ing a neural code, Science 252: 1854-1857. Bower, J. M . and Beeman, D. (1995). The Book of Genesis, Springer-Verlag. Brooks, R. A. (1986a). Achieving artificial intelligence through building robots, A.I. Memo 899, M.I.T. Brooks, R. A. (1986b). A robust layered control system for a mobile robot, IEEE Journal of Robotics and Automation 2(1): 14—23. Campbell, S. and Wang, D. L. (1998). Relaxation oscillators with time delay cou-pling, Physica D 111: 151-178. CAN (1998). KVASER's controller area network (CAN) introduction. URL: (http: //www.kvaser.se/can/initial.htm)[accessed Apr 7, 1998]. Chalfie, M. , Sulston, J. E., White, J. G., Southgate, E., Thomson, J. N. and Brenner, S. (1985). The neural circuit for touch sensitivity in Caenorhabditis elegans, Journal of Neuroscience 5(4): 956-963. Chalfie, M. , Tu, Y., Euskirchen, G., Ward, W. and Prasher, D. (1994). Green fluo-rescent protein as a marker for gene expression, Science 263: 802-805. Chiel, H. J. and Beer, R. D. (1997). The brain has a body: Adptive behavior emerges from interactions of nervous system, body and environment., Trends in Neu-roscience 20: 553-557. Cliff, D. (1991). Computational neuroethology: A provisional manifesto, in J. A. Meyer and S. W. Wilson (eds), From Animals to Animats: Proceedings of the First International Conference on Simulation of Adaptive Behaviour, MIT Press, pp. 29-39. Comer, D. E. (1995). Internetworking with TCP/IP, Volume I, Third Ed.: Princi-ples, Protocols, and Architecture, Prentice-Hall. Comer, D. E. and Stevens, D. L. (1993). Internetworking with TCP/IP, Volume III, BSD Socket Version: Client-Server Programming and Applications, Prentice-Hall. 110 Corey, D. P. and Garcia-Anoveros, J. (1996). Mechanosensation and the DEG/ENaC ion channels, Science 273: 323-324. Crochiere, R. and Rabiner, L. (1983). Multirate Digital Signal Processing, Prentice-Hall. Davis, R. E. and Stretton, A. O. W. (1989a). Passive membrane properties of mo-torneurons and their role in long-distance signaling in the nematode Ascaris, Journal of Neuroscience 9(2): 403-414. Davis, R. E. and Stretton, A. O. W. (1989b). Signaling properties of Ascaris mo-torneurons: Graded active responses, graded synaptic transmission, and tonic transmitter release, Journal of Neuroscience 9(2): 415^25. De Schutter, E., Angstadt, J. and Calabrese, R. (1993). A model of graded synaptic transmission for use in dynamic network simulations, Journal of Neurophysi-ology 69:1225-1235. Ermentrout, B. (1998). X-windows phaseplane plus auto. URL:http://mrb.niddk. nih.gov/xpp. Ermentrout, G. and Kopell, N. (1990). Oscillator death in systems of coupled neural oscillators, SIAMJ. Appl. Math. 50: 125-146. Funahashi, K. and Nakamura, Y. (1993). Approximation of dynamical systems by continuous time recurrent neural networks, Neural Networks 6: 801-806. Gear, C. W. (1971). Numerical Initial Value Problems in Ordinary Differential Equations, Prentice-Hall. Getting, P. (1989). A network oscillator underlying swimming in Tritonia, in J. Jacklet (ed.), Neuronal and Cellular Oscillators, Marcel Dekker Inc, New York, pp. 215-236. Gonzalez, J. and Tsien, R. (1995). Voltage sensing by fluorescence resonance en-ergy transfer in single cells, Biophys. J. 69: 1272-1280. Goodman, M . B., Hall, D. H., Avery, L. and Lockery, S. R. (1998). Active currents regulate sensitivity and dynamic range in C. elegans neurons, Neuron 20: 763-772. I l l Hall, D. H. and Russell, R. L. (1991). The posterior nervous system of the nema-tode Caenorhabditis elegans: serial reconstruction of identified neurons and complete pattern of synaptic interactions, Journal of Neuroscience 11(1): 1-22. Hammerstrom, D. (1995). Digital VLSI for neural networks, in M . A. Arbib (ed.), Handbook of Brain Theory and Neural Networks, MIT Press, pp. 304—309. Harris-Warrick, R., Nagy, F. and Nusbaum, M . (1992). Neuromodulation of stomatogastric networks by identified neurons and transmitters, in R. Harris-Warrick, E. Marder, A. Selverston and M . Moulins (eds), Dynamic Biological Networks: The Stomatogastric Nervous System, MIT Press, pp. 87-137. Harvey, I., Husbands, P., Cliff, D., Thompson, A. and Jakobi, N. (1997). Evo-lutionary robotics: The Sussex approach, Robotics and Autonomous Systems 20: 205-224. Hines, M . (1993). The NEURON simulation program., in J. Skrzypek (ed.), Neu-ral Network Simulation Environments, Kluwer Academic Publishers, Norwell, MA. Hopfield, J. (1995). Pattern recognition comutation using action potential timing for stimulus representation, Nature 376: 33-36. Hoppensteadt, F. and Izhikevich, E. (1997). Weakly Connected Neural Networks, Springer, New York. Horowitz, P. and Hill, W. (1989). The Art Of Electronics, second edn, Cambridge University Press. IEEE (1993). IEEE standard for information technology: Protocols for distributed interactive simulation applications. IEEE Std. 1278-1993. IEEE (1995). IEEE standard for a high performance serial bus. IEEE Std. 1394-1995. IEEE (1996). The POSIX system application program interface. IEEE Std. 1003.1-1988. 112 Jung, R., Kiemel, T. and Cohen, A. (1996). Dynamic behavior of a neural net-work model of locomotor control in the lamprey, Journal of Neurophysiology 75: 1074-1086. Kelso, J. A. S. (1995). Dynamic Patterns. The Self-Organization of Brain and Be-havior, MIT Press. Knuth, D. (1997). The Art Of Computer Programming, Vol 1: Fundamental Algo-rithm, third edn, Addison-Wesley. Koch, C. and Segev, I. (1989). Methods in Neuronal Modeling: From Synapses to Networks, MIT Press, Cambridge, MA. Lockery, S. R., Nowlan, S. J. and Sejnowski, T. J. (1992). Modelling chemotaxis in the nematode C. elegans, in J. Bower and F. Eechman (eds), Computation and Neural Systems, Kluwer Academic Publishers, Norwell, MA. Lockery, S. R. and Sejnowski, T. J. (1992). Distributed processing of sensory in-formation in the leech. III. a dynamical neural network model of the local bending reflex, Journal of Neuroscience 12(10): 3877-3895. Lygeros, J. (1996). Hierarchical Hybrid Control of Large Scale Systems, PhD the-sis, Department of Electrical Engineering, University of California, Berkeley. Mahowald, M . A. (1992). Evolving analog VLSI neurons, in T. McKenna, J. Davis and S. F. Zornetzer (eds), Single Neuron Computation, Academic Press. Marder, E. and Selverston, A. I. (1992). Modeling the stomatogastric nervous sys-tem, in R. M . Harris-Warrick, E. Marder, A. I. Selverston and M . Moulins (eds), Dynamic Biological Networks: The Stomatogastric Nervous System, MIT Press, pp. 161-196. Martin, F. (1995). Mini board 2.0 technical reference. URL: (http://fredm.www. media.mit.edu/people/fredm/papers/mb) [accessed Apr 27,1998]. Mills, D. (1990). On the accuracy and stability of clocks synchronized by the Net-work Time Protocol in the Internet system, ACM Computer Communication Review 20(1): 65-75. Mills, D. (1994). Precision synchronization of computer network clocks, ACM Computer Communication Review 24(2): 28-^ 1-3. 113 Mills, D. L. (1992). Network Time Protocol (version 3) specification, implementa-tion and analysis. IETF Network Working Group Report RFC-1305. Morse, T, Ferree, T. and Lockery, S. (1998). Robust spatial navigation in a robot inspired by chemotaxis in Caenorhabditis elegans, Adaptive Behavior 6: 393-410. Niebur, E. (1988). Theorie du systeme locomoteur et neuronal des Nematodes, PhD thesis, University of Lausanne. Niebur, E. and Erdos, P. (1991). Theory of the locomotion of nematodes: Dynamics of undulatory progression on a surface, Biophysical Journal 60: 1132-1146. Niebur, E. and Erdos, P. (1993). Theory of the locomotion of nematodes: Con-trol of the somatic motor neurons by interneurons, Mathematical Biosciences 118:51-82. Northmore, W. W. D. P. M . and Elias, J. G. (1997). Neuromorphic synapses for artificial dendrites, Analog Integrated Circuits and Signal Processing 13: 167— 184. Nyquist, H. (1928). Certain topics in telegraph transmission theory, AIEE Trans. pp.617-644. Oppenheim, A. V. and Schafer, R. W. (1989). Discrete-Time Signal Processing, Prentice Hall. Osborne, N. (1983). Dale's Principle and Communications Between Neurones, Pergamon Press, New York. Peak Audio Inc. (1998). CobraNet. URL: (http://www.peakaudio.com/public/ CobraNet/index.html) [accessed Apr 7,1998]. Pearlmutter, B. (1989). Learning state space trajectories in recurrent neural net-works, Neural Computation 1: 263-269. Pittsburgh Supercomputing Center (1998a). NEURON. URL: (http://www.psc.edu/ general/software/packages/neuron/neuron.html) [accessed Apr 7, 1998]. Pittsburgh Supercomputing Center (1998b). Parallel GENESIS at PSC. URL:(http: //www.psc.edu/general/software/packages/pgenesis/project_docs%) [ac-cessed Apr 7, 1998]. 114 Port, R. F. and van Gelder, T. (eds) (1995). Mind as Motion: Explorations in the Dynamics of Cognition, MIT Press. Press, W., Flannery, B., Teukolsky, S. and Vetterling, W. (1988). Numerical Recipes in C, Cambridge University Press, Cambridge. Profibus (1998). PROFIBUS. URL: (http://www.profibus.com) [accessed Apr 7, 1998]. Raizen, D. and Avery, L. (1994). Electrical activity and behavior in the pharynx of Caenorhabditis elegans, Neuron 12: 483-495. Rail, W. (1977). Core conductor theory and cable properties of neurons, in E. R. Kandel (ed.), Handbook of Physiology: The Nervous System, Vol. 1, American Physiological Society, Bethesda, MD. Rail, W. (1989). Cable theory for dendritic neurons, in C. Koch and I. Segev (eds), Methods in Neuronal Modelling: From Synapses to Networks, MIT Press, pp. 9-62. Rankin, C. H. (1991). Interactions between two antagonistic reflexes in the nema-tode Caenorhabditis elegans, J. Comp. Physiol. A 169: 59-67. Rankin, C. H., Beck, C. D. O. and Chiba, C. M . (1990). Caenorhabditis elegans: a new model system for the study of learning and memory, Behavioural Brain Research 37: 89-92. Real Networks, Inc (1998). Realtime streaming protocol resource center. URL: (http.7/www.real.com/rtsp) [accessed Apr 7,1998]. Real-Time Magazine (1998). Real-time encyclopaedia. URL: (http://www. realtime-info.be) [accessed Apr 7,1998]. Roehrig, C. J. and Rankin, C. H. (1998). Dymods I: A framework for modularizing dynamical neuronal structures, (submitted). Schinkmann, K. and Li , C. (1992). Localization of FMRFamide-like peptides in Caenorhabditis elegans, J. Comp. Neurol. 316(251-260). Schug, K. H. (1995). DIS NG - a flexible protocol for all simulation applications, Proc. 13th DIS Workshop on Standards for the Interoperability of Distributed Simulations, Orlando, Florida, Sep. 18-22. 115 Segev, I., Fleshman, J. W. and Burke, R. E. (1989). Compartmental models of com-plex neurons, in C. Koch, and I. Segev (eds), Methods in Neuronal Modelling: From Synapses to Networks, MIT Press, pp. 63-96. Skinner, F., Kopell, N. and Marder, E. (1994). Mechanisms for oscillation and frequency control in reciprocally inhibitory model neural networks, Journal of Computational Neuroscience 1: 69-87. Smith, J. O. and Gossett, P. (1984). A flexible sampling-rate conversion method, Proceedings of the IEEE Conference on Acoustics, Speech and Signal Pro-cessing, San Diego, Vol. 2, pp. 19.4.1-19.4.4. Stretton, A. O. W., Donmoyer, J. E., Davis, R. E., Meade, J., Cowden, C. and Sithigorngul, P. (1992). Motor behavior and motor nervous system function in the nematode Ascaris suum, J. Parasitol 78: 206-214. Strogatz, S. H. (1994). Nonlinear Dynamics and Chaos with Applications to Physics, Biology, Chemistry, and Engineering, Addison-Wesley. Tavernakis, N., Shreffler, W., Wang, S. and Driscoll, M . (1997). Unc-8, a DEG/ENaC family member, encodes a subunit of a candidate mechanically gated channel that modulates C. elegans locomotion, Neuron 18: 107-119. Thomas, J. H. (1990). Genetic analysis of defecation in Caenorhabditis elegans, Genetics 124: 855-872. White, J. G., Southgate, E., Thomson, J. N. and Brenner, S. (1976). The structure of the ventral cord of Caenorhabditis elegans, Philos. Trans. R. Soc. Lond. (Biol.) 275: 327-348. White, J. G., Southgate, E., Thomson, J. N. and Brenner, S. (1986). The structure of the nervous system of the nematode Caenorhabditis elegans, Philos. Trans. R. Soc. Lond. (Biol.) 314: 1-340. Wicks, S. R. and Rankin, C. H. (1995). Integration of mechanosensory stimuli in Caenorhabditis elegans, Journal of Neuroscience 15(3): 2434-2444. Wicks, S. R. and Rankin, C. H. (1997). The effects of tap withdrawal response habituation on other withdrawal behaviors: The localization of habituation, Behavioral Neuroscience 111: 342-353. 116 Wicks, S. R., Roehrig, C. J. and Rankin, C. H. (1996). A dynamic network simula-tion of the nematode tap withdrawal circuit: Predictions concerning synaptic function using behavioral criteria, Journal of Neuroscience 16(2): 4017-4031. Wood, W. B. (1988). The Nematode Caenorhabditis elegans, Cold Spring Harbor Laboratory, Cold Spring Harbor, NY. Yamaha (1998). mLAN: Audio and music data transmission protocol. URL: (http: //www.yamaha.co.jp/tech/1394mLAN/)[accessed Apr 7, 1998]. Yamauchi, M . M . and Beer, R. D. (1994). Sequential behavior and learning in evolved dynamical neural networks, Adaptive Behavior 2: 219-246. Yang, H. (1995). Intrinsic oscillations of neural networks, Proc. ICNN'95, Vol. 6, pp.3044-3047. Yang, H. and Dillon, T. (1994). Exponential stability and oscillation of hopfield graded response neural network, IEEE Trans, on Neural Networks 5: 719-729. Young Chang RDI (1998). Presto digital audio interface. URL: (http://www.ycrdi. com/presto) [accessed Apr 7, 1998]. 117 Appendix A The DSS Protocol Specification and API A . l The DSS Protocol Specification The DSS protocol implements layers 5 to 7 of the ISO 7-layer reference model (Comer, 1995) and relies on an underlying network transport protocol such as TCP/IP or IEEE 1394. The protocol has three main components (Figure A.l) . The DSS Signal component deals with messages that carry the signal data. This component is re-sponsible for presenting a continuous signal abstraction to the receiver application, and performs the necessary signal reconstruction and resampling as required by the receiver. It uses an unacknowledged isochronous datagram transport service. The DSS Control component is responsible for establishing, managing and break-ing DSS connections. It uses an acknowledged datagram transport service. The 118 DSS API DSS , r protocol , DSS Signal DSS Control DSS Name Service ISOC Transport ASYNC Transport > Network Transport Figure A . l : DSS Architecture DSS Name Service provides a name registry to allow DSS signal connections to be specified by a symbolic name rather than numerical addresses. Much of the nomenclature used here has been adopted from the TCP/IP protocol (Comer, 1995). An octet refers to an 8-bit quantity (i.e. a byte on most computers). All numerical quantities appearing in DSS messages are in network byte ordering (i.e. MSB first), and are aligned on type boundaries. A. l . l DSS Port Addresses DSS port addresses have two components: a 32-bit local address to uniquely identify the port on a single host computer or device, and an 8-octet transport ad-dress to uniquely identify the host. The transport address consists of a single octet to identify the transport family, plus a 7-octet transport-dependent address (see Fig-ure A.2). The transport family is 1 for TCP/IP using UDP, and 2 for IEEE 1394. 0 8 16 24 31 local DSS port transport tamily transport-dependent address transport-dependent address Figure A.2: DSS Address 119 In the TCP/IP UDP implementation of DSS version 1, the 7-octet transport address octets are in order: a zero-pad octet, a 2-octet UDP port number in network byte order, and a 4-octet IP address in network byte order. A.1.2 DSS Messages DSS messages are an integral number of quadlets (4-octets) and consist of a DSS header followed by the message contents. DSS messages can be one of two types: isochronous (ISOC) or asynchronous (ASYNC). ISOC messages are unacknowledged and used for carrying the time-critical signal data. ASYNC mes-sages are used for control and name service functions and are acknowledged using a sequence number to match acknowledgement responses to requests. To allow for efficient embedded implementations and mappings onto transport layers such as IEEE-1394, all DSS messages are guaranteed to have a size of 200 octets or less. The two message types were designed to be efficiently mapped onto the asyn-chronous and isochronous services provided by the underlying transport layer. In the case of IEEE-1394, these would be the similarly named services. DSS version 1 was implemented on TCP/IP using UDP datagrams to provide both services. The DSS Header The DSS message header is shown in Figure A.3. The four high-order bits of the header is used to identify the version of the DSS protocol, encoded in reverse bit order. The current version is 1 and it is en-coded as 0x8. Bits 12-15 of the first header quadlet contain message flags as shown 120 VERS 8 reserved 16 FLAGS 24 31 MSGTYPE MSGLEN DSS src port DSS dst port in Table A. 1: flag Figure A.3: DSS Message Header value description DSS_ACKREQ 0x1 requires an a acknowledgement DSS_ACK 0x2 this is an acknowledgement DSS_TIMEOUT 0x4 request timed out D S S _ F A I L 0x8 request failed Table A . l : DSS Message Header Flags The MSGTYPE octet encodes the DSS message type as described below. The MSGLEN octet encodes the size of the DSS message contents in octets, not including the DSS header. The DSS source port and destination port identify the endpoints of the DSS connection. To form a complete DSS address, they must be combined with the transport-dependent address which must be obtained from the transport layer. Since DSS messages are encapsulated in transport-layer datagrams, these datagrams must include the source transport address if the full DSS source address is to be recoverable. This is the case for TCP/IP (Comer, 1995) and IEEE-1394 (IEEE, 1995). I S O C Message ( M S G T Y P E = 1 ) An isochronous message (Figure A.4) carries signal data. The sample times-tamp is the number of microseconds since the connection epoch. In DSS version 1, the signal data is a single sample encoded as an integer scaled up by a factor of 121 16 24 3J signal tlmestamp signal data Figure A.4: DSS Isochronous Message 106. This allows a representation of signal values in the range -2147 to 2147 with a precision of 10 - 6 . Future DSS versions will specify the signal and timestamp format at connection time and may encode multiple samples in a single message. A S Y N C Message (MSGTYPE=2) A generic asynchronous message is shown in Figure A.5. This message is 0 8 16 24 31 SEQNO reserved Figure A.5: DSS Asynchronous Message used for simple acknowledgments and to "ping" ports to see if they are alive. For acknowledgements, the DSS_ACK bit is set in the header and SEQNO identifies the original message that requested the acknowledgment. For pings, the DSS_-ACKREQ bit is set in the header flags, and the destination port returns an ASYNC acknowledgement message. A 16-bit data field can be used to return error codes or other data. C O N N E C T Message (MSGTYPE=3) A connection request message is shown in Figure A.6. This message is sent by an output port to a target input port to request its connection. The DSS source address is the originating output port of the signal. FORMAT indicates the 122 8 SEQNO 16 24 FORMAT 31 DSS source address DSS source address DSS source address SRATE EPOCH Figure A.6: DSS Connection Request Message data format for the signal and is currently unused in DSS version 1. In future versions, this will indicate the encoding format and scaling factor for the signal data. SRATE is the nominal sampling rate (i.e. twice the bandwidth) of the signal in thousandths of Hertz. SRATE is used to determine the filter cutoff frequency of the reconstruction filter. EPOCH is the reference time for the connected signal's timestamps in seconds since midnight, January 1, 1970, GMT. This value is added to the signal's timestamps to give their actual absolute time reference. An ASYNC acknowledgement packet is returned to the sender upon completion. DISCONNECT Message (MSGTYPE=4) The DISCONNECT message (Figure A.7) is sent to a target DSS input port by an output port to request its disconnection. The DSS source address is the orig-8 16 24 31 SEQNO reserved DSS source address DSS source address DSS source address Figure A.7: DSS Disconnection Request Message inating output port of the signal. An ASYNC acknowledgement packet is returned to the sender upon completion. 123 A D D T A R G E T Message (MSGTYPE=5) The ADD TARGET message (Figure A.8) is sent to a DSS output port to request its connection to a target input port. The DSS destination address is the 8 16 SEQNO 24 31 reserved DSS target address DSS target address DSS target address Figure A.8: DSS Add Target Message address of the target input port to be connected. An ASYNC acknowledgement packet is returned to the sender upon completion. If the output port was unable to contact the target port, the D S S _ T I M E O U T bit is set in the acknowledgement packet. D E L T A R G E T Message (MSGTYPE=6) The DEL TARGET message (Figure A.9) is sent to a DSS output port to re-quest its disconnection from a target input port. The DSS destination address is the 8 16 SEQNO 24 31 reserved DSS target address DSS target address DSS target address Figure A.9: DSS Delete Target Message address of the target input port to be disconnected. An ASYNC acknowledgement packet is returned to the sender upon completion. If the output port was unable 124 to contact the target port, the D S S _ T l M E O U T bit is set in the acknowledgement packet. N A M E REGISTER Message (MSGTYPE=7) The NAME REGISTER message (Figure A. 10) is sent to a DSS name server to register the name with a DSS address. NAMELEN is the length in octets of the 0 8 16 24 31 SEQNO NAMELEN reserved DSS target address DSS target address DSS target address DSS name string Figure A. 10: DSS Name Register Message name string to be registered, not counting any padding octets. DSS address is the address to be registered with the name. DSS name string is a sequence of ASCII octets, padded to a quadlet boundary. An ASYNC acknowledgement packet is returned to the sender upon completion. N A M E QUERY Message (MSGTYPE=8) The NAME QUERY message (Figure A . l 1) is sent to a DSS name server to lookup a DSS address by name. NAMELEN is the length in octets of the name 0 8 16 24 31 SEQNO NAMELEN reserved DSS name string Figure A . l l : DSS Name Query Message 125 string to be looked up, not counting any padding octets. DSS name string is a sequence of ASCII octets, padded to a quadlet boundary. A N A M E RESPONSE acknowledgement packet is returned is returned to the sender upon completion. If the lookup was unsuccessful, the DSS_FAIL flag is set in the header of the N A M E RESPONSE packet. NAME RESPONSE Message (MSGTYPE=9) The N A M E RESPONSE message (Figure A. 12) is an acknowledgement packet containing the response to a N A M E QUERY message. SEQNO identifies 1£_ SEQNO 31 reserved DSS DSS address DSS address Figure A. 12: DSS Name Response Query the original N A M E QUERY message that requested the name lookup. If the DSS_-FAIL flag is not set, the DSS address is the address that is registered with the name. EPOCH Message (MSGTYPE=10) The EPOCH message (Figure A. 13) is used to renegotiate the connection's epoch as described below. EPOCH is encoded in the same format as in the CON-0 8 16 24 31 SEQNO reserved EPOCH Figure A. 13: DSS Epoch Query 126 N E C T message. A n A S Y N C acknowledgement message is returned to the sender upon completion. The epoch renegotiation proceeds as follows. Let E0 be the current connec-tion epoch, and let M A X _ T I M E S T A M P be the maximum timestamp interval encoded in 32 bits (this is 2 3 2 microseconds in DSS version 1, but in future versions may de-pend on the encoding format). The relative timestamp is difference between the absolute timestamp and the current connection epoch. When the time is such that the relative timestamp exceeds M A X T I M E S T A M P / 2 , the connection is in an epoch renegotiation phase. Timestamps relative to the old epoch E0 can be identified by a 1 in the high-order bit. A new epoch E\ is established by the signal output port and sent to the input port which records and acknowledges it. When the output port has received the acknowledgement, it begins using the new E\ epoch for its timestamp references. Figure A . 14 depicts the process. This mechanism ensures that any out-of-order samples w i l l still have their timestamps computed correctly. This method assumes that epoch renegotiation takes less than M A X T I M E S T A M P / 2 and that packet transmission delays are less than M A X _ T I M E S T A M P . A.2 The DSS Application Programmer Interface (API) The DSS application programmer interface (API) is a set of C program functions that provides an application with the capability to interact with DSS sig-nals. The interface abstraction is similar to the B S D sockets interface (Comer and Stevens, 1993) and consists of the following functions. 127 DSS OUTPUT Port EPOCH RENEGOTIATION PHASE EPOCH ACK I . E . DSS INPUT Port l i - E . < MAX_TIMESTAMF/2 l 2 • K0 > HAXilTIMESTAMP/2•: EPOCH ACK t} - E„ | > MAX_TIMESTAMP / 2 t 4 - E , < HAX_TIMESTAMP/2 Figure A. 14: Epoch renegotiation. The boxes represent DSS messages sent between a DSS output port and a DSS input port and time flows downwards. The dashed ar-rows show isochronous messages carrying signal data, and the solid arrows show the control messages involved in epoch renegotiation. The relative timestamp con-tained in an isochronous message is shown as tt — En where is the absolute timestamp of that sample datum, and En is the epoch in effect when that message was sent. The shaded area represents the epoch negotiation phase (see text). void dss_init( char *nameserver ) Initializes the DSS subsystem. The nameserver is the transport-dependent name string that identifies the host functioning as the DSS name registry, void dss_fini() Terminates the DSS subsystem, dss socket dss open input( char *name, double rate) Creates an input socket called name, registers it with the name server and returns it. The rate is the expected rate at which samples will be read from the socket and is used to determine the interpolation filter characteristics, dss socket dss open output( char *name, double rate, int format) Creates an output socket called name, registers it with the name server and returns it. The rate is the expected rate at which samples will be written to 128 the socket and is used to determine the interpolation filter characteristics of the receiver. In DSS version I, format is unused; in future versions it will specify the signal data format, void dss_close( dss socket s) Closes socket s. Disconnects all open connections and removes its name from the registry. int dss_lookup( char *name, struct dss_addr *addr ) Looks up the name in the DSS registry and returns its full DSS address in addr. Returns 1 if successful, 0 otherwise, int dss_ping( struct dssaddr *addr ) Pings a DSS port to see if it is alive. Returns 1 if successful; 0 if the ping times out. double dss_read( dss_socket s, double t) Interpolates and returns the signal's value from input socket s for the time t. This function performs the signal reconstruction from the currently available samples and will therefore only be accurately reconstructed for t values in the interval [t, t + dt] where dt is the sampling interval of the received signal. Therefore, it should be scheduled in the following manner: while ( wallclockO < t ) usleep ( S L E E P _ I N T E R V A L ) ; /* wait */ x = dss_read( s, t ); where SLEEP_INTERVAL is a suitably small fraction of dt. If the system supports better real-time scheduling facilities, they should be used instead of usleep. 129 int dss getinfo( struct dssinfo *info ) Gets information associated with input socket s: struct dss_info{ double delay; int bufsize; int samples; The de l ay is established at connection time and is determined by the inter-polation filter width. The b u f s i z e is the size of the signal input buffer and samples is the number of samples used in the most recent interpolation (in-vocation of d s s r e a d ) . int dss addtarget( dss_socket s, struct dss_addr *addr) Adds the DSS port with address addr to the list of targets for output socket s. Returns 1 if successful; 0 otherwise, int dss_deltarget( dss_socket s, struct dssaddr *addr) Removes the DSS port with address addr from the list of targets for output socket s. Returns 1 if successful; 0 otherwise, void dss_write( dss_socket s, double t, double val) Writes the signal's value val to the output socket ^ with timestamp t. The caller is responsible for ensuring that the sampling interval is sufficient to rep-resent the signal's bandwidth as declared in the rate parameter of the d s s -open_output call, and for scheduling calls to d s s_wr i te within the in-terval [t, t + dt] where dt is the sampling interval. Otherwise, the receiver may not receive the necessary samples for accurate reconstruction. In DSS version 1, dss_addtarget operates on an open output socket. In future versions, the d s s a d d t a r g e t function will take two DSS addresses and will not require an open socket. This will allow connections to be established and managed remotely by a connection manager. The following example illustrates how to set up a DSS output port to deliver a signal computed by f (t) to a destination DSS input port with name l i s t e n e r . The sampling interval is dt. d s s _ a d d r d s t ; d s s _ s o c k e t * m e ; d s s _ i n i t ( ) ; me = d s s _ o p e n _ o u t p u t ( " s e n d e r " , 1 . 0 / d t , 0) ; i f ( d s s _ l o o k u p ( " l i s t e n e r " , & d s t ) ) d s s _ a d d t a r g e t ( m e , & d s t ) ; f o r ( t = w a l l c l o c k ( ) . ; t < M A X _ T ; t + = d t ) { w h i l e ( w a l l c l o c k O < t ) u s l e e p ( S L E E P _ I N T E R V A L ) ; / * w a i t * / d s s _ w r i t e ( m e , t , f ( t ) ) ; } The following example illustrates how to set up a DSS input port that reads a signal and compares it to a computed version. Its sampling interval is dt , which is unrelated to the sampling interval of the transmitted signal. d s s _ s o c k e t * m e ; d s s _ i n i t ( ) ; me = d s s _ o p e n _ i n p u t ( " l i s t e n e r " , 1 . 0 / d t ) ; d e l a y = d s s _ d e l a y ( m e ) ; f o r ( t = w a l l c l o c k ( ) ; t < M A X _ T ; t + = d t ) { w h i l e ( w a l l c l o c k O < t ) u s l e e p ( S L E E P _ I N T E R V A L ) ; / * w a i t * / x = d s s _ r e a d ( m e , t ) ; p r i n t f ( " d i f f e r e n c e = % g \ n " , x - f ( t - d e l a y ) ) ; 131 Appendix B Robot Assembly Instructions and mblibAPI B.l LEGO Tank Robot Construction The assembly diagrams for the the LEGO tank robot of Chapter 5 are shown in Figure B . l - Figure B.5. 132 P a g e 1 Figure B . l : Robot Tank Chassis LEGO Assembly, Part 1 133 P a g e 2 Figure B.2: Robot Tank Chassis LEGO Assembly, Part 2 134 Insert battery clips in slots before attaching top 8x1 battery cover beams. "| 7 Insert a large 40-tooth gear into the slot and impale it with short 3cm axle. Attach a small 8-tooth gear to the other end of the ax Attach a large 40-tooth gear to a 5 cm axle and insert it so that it meshes with the 8-tooth gear. Secure it on the other end with an axle "nut". P a g e 3 Figure B.3: Robot Tank Chassis LEGO Assembly, Part 3 135 Page 4 Figure B.4: Robot Tank Chassis LEGO Assembly, Part 4 136 20 Sprocket Positioning Attach freewheeling sprockets at indicated positions. You'll need to use snap-in axles for the middle pair of small sprockets because there is no clearance on the other side of the hole. Snap-in axles aren't found in kit 9605, but can be found in other Technics kits. With a little modification, the two rear-most sprockets can be mounted on shaft-encoders. The rear engine space was designed to accommodate a modified ALPS "press brush contact" type rotary encoder from an old (circa 1990) Microsoft mouse. 24-tooth sprockets 40-tooth sprocket Link together 74 chain links to form each tread. P a g e 5 Figure B.5: Robot Tank Chassis LEGO Assembly, Part 5 137 B.2 Miniboard Host C Library The m b l i b host library consists of the following API: m b o p e n ( p o r t ) Open and return a miniboard descriptor using the given serial port, mb c l o s e ( m b ) Close the given miniboard descriptor, mb c l e a n u p ( ) Clean up the miniboard system after an error, mb r u n ( m b , i n t e r v a l , u s e r f u n c , u s e r p a r m Start periodically updating the Miniboard at the specified interval. Af-ter each update, the user supplied function is called with the provided parameter, mb s t o p ( m b ) Stops the periodic Miniboard updating, m b s e t m o t o r p w r (mb, n , o n ) Turns motor n on or off. m b m o t o r p w r (mb, n ) Returns the on/off state of motor n. m b _ m o t o r _ s e t _ d i r (mb, n , d i r ) Sets the direction for motor n. m b _ m o t o r _ d i r (mb, n ) Returns the direction for motor n. mb m o t o r s e t s p e e d ( m b , n , s p e e d ) Sets the speed for motor n. mb m o t o r s p e e d ( m b , n Returns the speed for motor n. 138 m b s e t d i g i t a l d i r ( m b , n , d i r ) Sets the direction for digital I/O bit n. mb d i g i t a l d i r ( m b , n ) Returns the direction for digital I/O bit n. mb s e t d i g i t a l ( m b , n , v a l u e ) Sets the output value for digital I/O bit n. mb d i g i t a l ( m b , n ) Returns the state of digital I/O bit n. mb s w i t c h ( m b , n ) Returns the complement of the state of digital I/O bit n (more intuitive when using switches). m b _ a n a l o g ( m b , n ) Returns the value of analog input n. m b _ d s h a f t ( m b , n ) Returns the number of milliseconds between edges on the digital shaft n. mb r e a d v e r s i o n (mb) Reads the BINMON version string, mb r e a d b y t e (mb, a d d r ) Reads the byte at Miniboard address addr. m b w r i t e b y t e (mb, a d d r , d a t a ) Writes the given byte of data at Miniboard address addr. mb r e a d w o r d ( m b , a d d r ) Reads the 16-bit word at Miniboard address addr. mb w r i t e w o r d ( m b , a d d r , d a t a ) Writes the given word of data at Miniboard address addr. 139 mb r e s e t ( m b ) Resets the Miniboard. For a full description of this API, see the m i n i b o a r d . h file in the m b l i b pack-age. 140 


Citation Scheme:


Citations by CSL (citeproc-js)

Usage Statistics



Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            async >
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:


Related Items