運用同化與調適於多代理人的合作學習

Applying Assimilation and Accommodation for Cooperative Learning of Multi-Agent

盧志偉(Chih-Wei Lu)

指導教授 ： 郭忠義

國立臺北科技大學/電資學院/資訊工程系研究所/碩士(2010年)

https://doi.org/10.6841/NTUT.2010.00394

摘要

本篇論文結合皮亞傑同化調適的概念，以及強效式學習和倒傳遞類神經網路技術，發展代理人適應動態多變環境的知識與能力。智慧型代理人面對不確定的環境資訊，根據舊有的知識與能力，對自己的知識架構進行同化或調適，使其能更適應環境的變化。我們以Robocup足球模擬平台進行足球競賽，以驗證本研究所提出的方法。

關鍵字

RoboCup ； 同化 ； 調適 ； Q-Learning

並列摘要

Adapting learning is the essential ability to improve the convergence rate and learning quality in the multi-agent system. A cooperative mechanism needs learning cooperatively between agents. This research applies assimilation and accommodation in complex environment to product effective action. There are intentional schema and perceptional schema in our assimilation and accommodation. In intentional schema, reinforcement learning is used to choose target state. In perceptional schema, back-propagation neural network is used to predict environmental forward dynamics. When the error between predicting state and actual state is too large, it means our knowledge can’t assimilate this sample. So we must adjust our knowledge to fit it. This is an accommodation process. We use the RoboCup simulator to explain our research.

並列關鍵字

RoboCup ； Assimilation ； Accommodation ； Q-Learning ； Neural Network

參考文獻

[1] A. Rao, M. Georgeff, “BDI Agents: From Theory to Practice”, International Conference on Multi-Agent Systems, 1995.

[2] Marko Verbeek. “3APL as programming language for cognitive robots”, Masters' Thesis Technical Artificial Intelligence Computer Science, 2002.

[3] E.C ten Hoeve, “3APL Platform”, Master’s thesis Computer Science, 2003.

[4] J. Y. Kuo, M. L. Tsai, and N. L. Hsueh, “Goal Evolution based on Adaptive Q-learning for Intelligent Agent”, IEEE International Conference on Systems, 2006.

[8] L. Waltman, U. Kaymak, “A Theoretical Analysis of Cooperative Behavior in Multi-agent Q-learning”, IEEE International Symposium on Approximate Dynamic Programming and Reinforcement Learning, pp. 84-91, 2007.