Reinforcement Learning with conditioned role updating to prevent conflicts during the allocation of tasks (コンカレント工学)

概要

論文の詳細を見る
In many applications, effective task-allocation and task-completion strategies are crucial to achieve an optimal performance. Reinforcement Learning (RL) methods have been proposed for those applications in which the model of the system is unknown. However, RL methods usually aim for agents to learn behavior policies for the completion of a task, and rarely focus on the task-allocation problem. Furthermore, only few of the RL methods addressing the task allocation problem consider the possibility of conflicts among agents during the allocation of tasks. In this paper, we turn the task-allocation problem into a task-response problem, and propose a learning algorithm, implementing conditioned rule updating, to make agents learn conflict-free task-response policies. We use a single-shaft multi-car elevator (MCE) system of two cars as test application. Through simulations, we analyze the learning performance of our algorithm and its effectiveness in preventing interference events in the MCE system. Simulation results show that the conditioned rule-updating feature of our method improves its convergence speed and its effectiveness in searching for optimal policies. In addition, the results also demonstrate that with our algorithm elevator agents can learn task-response policies that allow them to service calls without interfering with each other, under an interfloor traffic pattern.
2011-01-13