代理人自我覺察能力對於個體積效與合作行為之研究

Abstract

本研究從智慧型代理人的角度出發，回應傳統人工智慧領域中機器學習的核心問題：以世界模型為基礎的學習型代理人的設計不足之處。為了使代理人具備自我學習的能力，換言之，類似人類的自我覺察能力，我們提出一套新穎的代理人認知學習架構，包含外在學習與內在認知雙重概念，並相容於舊有的代理人架構。同時，為了有效驗證此架構是強固的、可靠的，可廣泛地應用在許多實際的狀況，如電子商務環境、社會科學模擬系統，我們將代理人的目標與生存環境之間所造成的衝突，以反覆囚犯困局來模擬與實驗，透過代理人的人格特質分析，我們提出一套以超我層次為自我覺察目標的代理人自覺模型，並以前述的反覆囚犯困局的實驗結果，來分析自我覺察能力對於學習型代理人的個體績效與合作行為的影響。

根據實驗結果，本研究證明以超我層次為自我覺察目標的代理人自覺模型，可有效幫助學習型代理人提升表現成效，並使合作行為提前浮現。更進一步的模擬與分析，我們發現只需要少數的代理人具備自覺能力，即可提升整體代理人社會公益，這些實驗結果也同時驗證本研究提出的代理人認知學習架構。最後，我們期望本研究所探討的方向與內容能讓大家重新思考與重視自我覺察對學習型代理人設計的重要性。

The approach of this research, how intelligent agents learn, is to deal with a core problem of Machine Learning. The problem of traditional artificial intelligence lies in the flaw that learning agents are designed on the basis of World Model. To endue agents with Self-Learning ability, in other words, the ability similar to self-awareness of human beings, we proposed a new cognitive learning model, which includes both external learning and internal cognition compatible with the former structure, for agents, called Agent Cognition Learning Model (ACLM). In order to prove this model is robust, reliable, and extensively applicable for real situations such as E-commerce or social science simulation systems, we will simulate and experiment the conflict between the societal and self-interested goals of agents with Iterative Prisoner’s Dilemma on Social Networks. Through an analysis on personalities of agents, we proposed a Self-Awareness Model in Superego Level for agents. With the experimental results, we will analyze how individual performance and collaborative behavior of learning agents would be affected.

The results of the experiments would demonstrate that the self-awareness model aim for superego level could certainly improve the performance of learning agents and expedite the emergence of collaborative behavior. Further simulation and analysis would show that as few of the agents are capable of self-awareness, the whole social benefits of agents would be enriched. These results also very strong the agent cognitive learning model proposed by our research. Finally, we hope this research could make people reconsider the importance of self-awareness in the design of self-learning agents.