3208 方策の初期値をニューラルネットに類推させる強化学習に関する研究(J25-1 メカニカルシステムとその知能化(1),J25 メカニカルシステムとその知能化)

概要

論文の詳細を見る
By clustering the encountered states by applying Fuzzy-ART to the state-space segmentation, it was possible to prevent an explosive increase of category units in state space and improve learning speed. Even if we used Fuzzy-ART, it was necessary to learn an action from the beginning when we encountered unknown states. We supposed that we were able to act appropriately for the unknown states, if we were able to extract common information of states to which a similar action was required. We proposed a reinforcement learning method with a neural-network which studied the relations between states and actions obtained by the reinforcement learning. The neural-network decided the initial value of policy for unknown states. We confirmed that this method improved the learning speed.
社団法人日本機械学会の論文
2007-09-07