Open Collections

Making predictions directly from past experiences

University of British Columbia

This thesis considers the problem of making predictions about new experiences based upon past experiences. The problem is of interest to artificial intelligence because past experiences are a kind of domain knowledge that is readily available to com- putational agents, and are at least one form of knowledge that humans use to make predictions. Instead of considering the problem in terms of first inducing a domain model from a set of past experiences, and then using some form of deduction to make predictions, this thesis develops a new technique called the reference class approach (RCA) that directly infers estimates of conditional probabilities from a knowledge base of past experiences. The resulting estimates can be readily used in a number of contexts such as non-monotonic reasoning, the characterisation of probability distribution functions, prediction and classification. Given a knowledge base (KB) of descriptions of past experiences, a description of a new experience, and a proposition representing a query about the new experience, the RCA estimates the conditional probability of the proposition being true of the new experience. The RCA starts by identifying a subset of the KB called the reference class that contains all those past experiences in the KB whose descriptions cover everything that is known about the new experience in addition to providing a truth value for the proposition. If there are no directly applicable past experiences, i.e., the reference class is empty, then the description of the new experience is modified until a non-empty ref- erence class can be found. This thesis investigates two new approaches to modifying the description, namely syntactic generalisation and chaining. Previous research has proposed that logical implication can be used to semantically generalise an empty reference class to any non-empty reference class. This thesis shows that semantic generalisation does not work in the context of making predictions from a KB of past experiences. This thesis argues that we should syntactically generalise the descrip- tion of the new experience. Chaining is a novel extension of syntactic generalisation that allows us to systematically increase what we know about a new experience by elaborating its description while generalising. Once a non-empty reference class has been identified the RCA estimates the conditional probability of the proposition being true by measuring the frequency with which the proposition is true in the reference class. The RCA is an inductive technique in that it estimates probabilities directly from past experiences. One useful test of an inductive technique is to test whether or not it can be used to make accurate predictions from past experiences. This thesis argues that in order to implement the RCA we need a notion of irrelevance to pick the most appropriate generalised or chained reference class. This thesis shows that even with very simple notions of irrelevance, the RCA’s estimates can be used to make predictions whose accuracy compares favourably with state of the art machine learning techniques on standard test data from the machine learning community.

Thesis/Dissertation

application/pdf

2009-04-09

For non-commercial purposes only, such as research, private study and education. Additional conditions apply, see Terms of Use https://open.library.ubc.ca/terms_of_use.

http://hdl.handle.net/2429/7008

Computer Science

University of British Columbia

Graduate