UBC Theses and Dissertations
Resource allocation optimization of a disrupted interdependent system using machine learning Khouj, Mohammed Talat
National safety and homeland security of an urban community rely heavily on a system of interconnected critical infrastructures (CI’s). The interdependencies in such complex systems give rise to vulnerabilities, which must be accounted for by a proper disaster management. It is a proactive step that is needed to address and mitigate any major interruption in a timely manner. Only then will the management of CI’s be able to appropriately reallocate and distribute the available scarce resources of an existing interdependent system. In this research, we propose an intelligent decision making system that optimizes the allocation of the available resources following an infrastructure disruption. The novelty of our suggested model is based upon the application of a well-known Machine Learning (ML) technique called Reinforcement Learning (RL). This learning method is capable of using experience from a massive number of simulations to discover underlying statistical regularities. Two alternative approaches to intelligent decision-making are studied, learning by Temporal Differences (TD) and Monte Carlo (MC) based estimation. The learning paradigms are explored within the context of competing designs composed of simulators and learning agents architected either independently or together. The results indicate the best learning performance is obtained using MC within a homogeneous system. The goal here is to maximize the number of discharged patients from emergency units by intelligently utilizing the existing limited resources. We show that such a learning agent, through interactions with an environment of simulated catastrophic scenarios (i2Sim-infrastrucutre interdependency simulator), is capable of making informed decisions in a reasonable time.
Item Citations and Data
Attribution-NonCommercial-NoDerivs 2.5 Canada