Study on Emergent Intelligence from Intelligent Agent Team under Dynamic Environments.

本研究对“动态环境下智能智能体团队的涌现智能”进行了全面的研究和实践，从理论构建到原型开发都得到了全面的检验。(1)以商用电动轮椅的改装为主要主题，以“智能轮椅”为主要研究对象，进行了原型研究。在原型轮椅的规范下，我们成功地从智能体团队的交互中产生了以自主、协作和协同行为为核心的智能行为。(2)围绕群体行为，利用智能智能体对宏观行为进行了基础性研究。人群行为和交通流涌现模拟实验的结果表明，群体行为和交通流的自主组织现象和极其复杂的行为导致了严重的事故。(3)在模拟或构造性方法的每一种情况下，为了在动态环境中获得高效的行为，智能体需要附加自适应学习和进化方法。发展了先进的强化学习和进化算法技术，提出了在智能轮椅控制和多智能体仿真中使智能体获得智能行为的实用技术。(4)为使智能体在故障情况下能够自主恢复，开发了一种具有进化硬件的动态配置系统，这是近年来的一个显著趋势。仿真实验的结果表明了该方法在解决自主机器人迷宫问题上的优势。作为这些研究的结果，我们证实了该项目对基础技术的发展和多智能体应用的实际应用的完成都具有重要的贡献。

With comprehensive studies and practical conclusions for "emergent intelligence from intelligent agent team under dynamic environments," this study has been comprehensively examined from a theory construction to a prototype development.(1)"Intelligent wheelchair," the major theme of this project, has been prototyped with converting commercial electric wheelchair. Under specifications of the prototyped wheelchair, we have succeeded in emergence of intelligent behavior focused on autonomous, cooperative, and collaborative actions from interaction in agent team.(2)Focusing on "crowd behavior", we conducted fundamental studies of macro-level behavior with intelligent agents. The results of emergence simulation experiments of crowd behavior and traffic flow have shown the phenomenon of autonomous organization and extremely complex behavior which leads to serious accidents.(3)In each case of simulation or constructive approaches, in order to acquire efficient behavior in dynamic environment, agents need to attach the adaptive learning and evolutionary approach. Developing advanced technique of reinforcement learning and evolutionary algorithm, we have proposed practical techniques which enable agents to acquire intelligent behavior in the intelligent wheelchair control and the niultiagent simulation.(4)A dynamic configuration system with "Evolutionary hardware," which is the notable trend of recent years, has been developed for enabling agent to have autonomous restoration in the case of fail. The result of the simulation experiment has shown the advantage at a maze problem of an autonomous robot.As the results of these studies, we confirmed that this project contributed importantly to the accomplishment of both development of fundamental techniques and showing practical applications for establishment of multiagent applications.