不完全知覚問題に対する内部メモリを用いた強化学習法に関する研究(セッション(6) : 一般)

概要

論文の詳細を見る
Perceptual aliasing problem is tackled by actor-critic reinforcement learning with internal memory. In this study, simple delayed response task that needs short-term memory ability is considered. Agent is given "cue stimuli" at first. Then, agent must maintain information of cue within delay period. Finally, "go sign" is given, and agent must respond correctly according to first cue. The relation between effectiveness of redundant memory and initial prediction of return is analyzed. Agent that has 3 features-redundant memory, optimistic initial prediction of return, greedy policy-shows relatively good performance. Ability of such agent is evaluated for several conditions of delay period length.
2004-03-08