Reinforcement learning models for transportation infrastructure management

Under the traditional approach to develop maintenance and repair policies for infrastructure facilities, Policy Evaluation and Policy Selection are performed while assuming that a complete and correct facility deterioration model is available. In addition, the time, cost and complexity required to develop such models are ignored in the framework. In order to address these limitations we formulate the problem of developing maintenance and repair policies as a Reinforcement Learning Problem. Under this approach, it is not necessary to model a facility's deterioration process to perform Policy Evaluation and Policy Selection. These functions are accomplished through mapping the effect of actions prescribed by the policies. In this paper, we explain the agent-system interaction considered in Reinforcement Learning. We discuss the probing-optimizing dichotomy that exists in the process of performing Policy Evaluation and Policy Selection. Then we describe Reinforcement Learning methods that can be used to address the problem of developing maintenance and repair policies. Finally, we present the results of a simulation study where we show the Reinforcement Learning is a viable approach that can be used to adapt and fine-tune policies in situations where complete and correct models of facility deterioration are not (yet) available.