A Rollout Method for Finite-Stage Event-Based Decision Processes

Event-based decision process (EDP) has provided a general framework for many event-based control, decision making, and optimization problems. Since the number of events could increase only linearly w.r.t. the system scale, EDP provides a computationally feasible way for many problems, which are time-consuming to solve in the Markov decision process (MDP) framework. Because the event sequence usually is not Markovian, policies that only depend on the current observable event usually are not optimal. In this paper, we develop a rollout method for finite-stage EDP, which uses simulation under a base policy to estimate the performances of action candidates. We show that this leads to a policy better than the base policy. This rollout method is easy to implement. The advantage is also demonstrated through a node activation policy optimization problem in wireless sensor network.