Reinforcement learning using monte carlo policy estimation for disaster mitigation

Abstract

Urban communities rely heavily on the system of interconnected critical infrastructures. The interdependencies in these complex systems give rise to vulnerabilities that must be considered in disaster mitigation planning. Only then will it be possible to address and mitigate major critical infrastructure disruptions in a timely manner. This paper describes an intelligent decision making system that optimizes the allocation of resources following an infrastructure disruption. The novelty of the approach arises from the application of Monte Carlo estimation for policy evaluation in reinforcement learning to draw on experiential knowledge gained from a massive number of simulations. This method enables a learning agent to explore and exploit the available trajectories, which lead to an optimum goal in a reasonable amount of time. The specific goal of the case study described in this paper is to maximize the number of patients discharged from two hospitals in the aftermath of an infrastructure disruption by intelligently utilizing the available resources. The results demonstrate that a learning agent, through interactions with an environment of simulated catastrophic scenarios, is capable of making informed decisions in a timely manner.