Accelerate Learning Processes by Avoiding Inappropriate Rules in Transfer Learning for Actor-Critic
スポンサーリンク
概要
- 論文の詳細を見る
This paper aims to accelerate processesof actor-critic method, which is one of majorreinforcement learning algorithms, by a transferlearning. In general, reinforcement learning is usedto solve optimization problems. Learning agentsacquire a policy to accomplish the target task autonomously.To solve the problems, agents requirelong learning processes for trial and error. Transferlearning is one of effective methods to acceleratelearning processes of machine learning algorithms.It accelerates learning processes by usingprior knowledge from a policy for a source task. Wepropose an effective transfer learning algorithm foractor-critic method. Two basic issues for the transferlearning are method to select an effective sourcepolicy and method to reuse without negative transfer.In this paper, we mainly discuss the latter. We proposedthe reuse method which based on the selectionmethod that uses the forbidden rule set. Forbiddenrule set is the set of rules that cause immediate failureof tasks. It is used to foresee similarity betweena source policy and the target policy. Agents shouldnot transfer the inappropriate rules in the selectedpolicy. In actor-critic, a policy is constructed by twoparameter sets: action preferences and state values.To avoid inappropriate rules, agents reuse only reliableaction preferences and state values that implypreferred actions. We perform simple experimentsto show the effectiveness of the proposed method. Inconclusion, the proposed method accelerates learningprocesses for the target tasks.
論文 | ランダム
- アウトブレイク発生時に現場で直接指導する (特集 永遠の課題を解決するワザ 手指衛生指導--これで私は結果を出した)
- 臨床・研究 出雲市における症候群サーベイランス
- 環境汚染が原因と考えられたMRSAアウトブレイクの2事例とICTの対応
- 胆汁由来Klebsiella oxytoca, Klebsiella pneumoniaeの分離頻度とこれらの菌の胆汁耐性に関する検討
- O1-008 抗インフルエンザ薬処方件数の増加から発見に至った院内インフルエンザアウトブレイク(一般演題 口頭発表,感染制御,医療薬学の創る未来 科学と臨床の融合)