強化学習によるSAの温度スケジューリング法の提案

概要

論文の詳細を見る
Simulated Annealing Method is used as an effective solution method of combinatorial optimization problems, but it has the problem that the settlement of the temperature which is an optimization parameter and scheduling are difficult. So, How to schedule temperature is proposed as an approach for this problem by using Q-Learning which is one of the reinforcement learning. With this technique, every time cooling is done, the state of the solution in a metropolis loop is observed, scheduling of temperature is performed from the state and the result of learning and it learns from that result again. By this technique, the loss in search is reduced by performing temperature scheduling using the history of search and even if a parameter is not adjusted in accordance with the problem, it is expected that it finds an answer firmly. From the result of this numerical experiment, the expected effect was acquired and the usefulness of this technique has been checked. A performance of the proposal methodology is discussed.
2004-12-08