Elucidation of communication emergence mechanism based on action time series in reinforcement learning agents.
| Project/Area Number |
25871049
|
|---|
| Research Category |
Grant-in-Aid for Young Scientists (B)
|
| Allocation Type | Multi-year Fund |
|---|
| Research Field |
Soft computing
Intelligent informatics
|
|---|
| Research Institution | Okinawa National College of Technology |
|---|
Principal Investigator |
Sato Takashi 沖縄工業高等専門学校, メディア情報工学科, 准教授 (70426576)
|
|---|
| Research Collaborator |
HASHIMOTO Takashi 北陸先端科学技術大学院大学, 知識科学系知識マネジメント領域, 教授 (90313709)
|
|---|
| Project Period (FY) |
2013-04-01 – 2017-03-31
|
| Project Status |
Completed (Fiscal Year 2016)
|
| Budget Amount *help |
¥4,550,000 (Direct Cost: ¥3,500,000、Indirect Cost: ¥1,050,000)
Fiscal Year 2015: ¥650,000 (Direct Cost: ¥500,000、Indirect Cost: ¥150,000)
Fiscal Year 2014: ¥2,990,000 (Direct Cost: ¥2,300,000、Indirect Cost: ¥690,000)
Fiscal Year 2013: ¥910,000 (Direct Cost: ¥700,000、Indirect Cost: ¥210,000)
|
|---|
| Keywords | ジェスチャー理論 / 原始的コミュニケーションの創発 / Q-learning / Neural Q-learning / Recurrent Q-learning / マルチエージェント・システム / 拡張版SOM / 暗示的フィードバック / 衝突回避ゲーム / 強化学習 / Recurrent-Q学習 / 基礎的行動の記号化 / Neural-Q学習 / ジェスチャー / Q学習 / マルチエージェント / コミュニケーションの創発 |
|---|
| Outline of Final Research Achievements |
Based on the gesture theory, we discussed an individual's ability and other factors necessary for emergence of proto-communication in a primitive society in which the communication was not established among the individuals. To verify the individual's ability aspect, we adopted a collision avoidance game and a reinforcement learning agents who can learn their action history as the game players. Our simulation showed that, by evaluating various models including a hybrid model between the Q-learning and the recurrent neural network, the abilities to learn and predict the past action history and its order can be played an important role in the emergence of communication. Also, to examine an element contributed to the formation of communication, we adopted a communication game with extended SOM learning agents. The second simulations suggested that "implicit feedback" obtained from situations other than individuals, which is proposed by us, can be improved the communication success rate.
|
Report
(5 results)
Research Products
(7 results)
