Learning and Evolution of Intelligent Systems Composed of Multi-individuals Interacting with Each Other

多个体交互组成的智能系统的学习与进化

基本信息

批准号：
14350212
负责人：
HIRASAWA Kotaro
金额：
$ 9.09万
依托单位：
Waseda University
依托单位国家：
日本
项目类别：
Grant-in-Aid for Scientific Research (B)
财政年份：
2002
资助国家：
日本
起止时间：
2002 至 2004
项目状态：
已结题

项目摘要

Intelligent system consisting of multi-individuals interacting between each other have been developed and the learning and evolution of the system have been studied. It is shown that the proposed system has better performance than conventional methods.1.Intelligent agents : Genetic Network Programming (GNP) has been developed as intelligent agents, which represents solutions using directed graph structures, and showed better performance than the conventional method in creating agent behavior.2.Interaction between agents : In order to create intelligent interaction between agents composed of GNP, Multi-Agent Systems with Symbiotic Learning and Evolution (Masbiole) has been developed, which applied symbiotic relations in the ecosystem such as Mutualism, Predation, Competition and Altruism. From the results of the simulations using Tileworld problem, it is clarified that the proposed method shows complex interaction between agents by introducing the symbiotic strategies.3.Learning and Evolution in Multi-agent systems : Generally, living things has been developing through evolution and learning. Evolution has been done for a long period of time, and learning is done based on trial-and-error during lifetime of each individual. Based on this idea, a GNP algorithm has been developed by -using evolution and reinforcement learning. This method can automatically create programs which show better performance than the GNP based on evolution only and the conventional evolutionary computation (Genetic Programming) because the proposed method can improve the programs during task execution (online learning) in addition to the evolution executed after task execution.

开发了由多个个体相互作用组成的智能系统，并研究了系统的学习和进化。结果表明，所提出的系统比传统方法具有更好的性能。 1.智能代理：遗传网络编程（GNP）已被开发为智能代理，它使用有向图结构表示解决方案，并且在创建方面比传统方法表现出更好的性能2.智能体之间的交互：为了在由GNP组成的智能体之间创建智能交互，开发了具有共生学习和进化的多智能体系统（Masbiole），该系统在生态系统中应用了共生关系，例如互惠、掠夺、竞争和利他主义。从使用 Tileworld 问题的模拟结果可以看出，该方法通过引入共生策略显示了智能体之间复杂的相互作用。3.多智能体系统中的学习和进化：一般来说，生物是通过进化和学习而发展的。进化已经进行了很长一段时间，学习是基于每个人一生中的反复试验而完成的。基于这个想法，利用进化和强化学习开发了 GNP 算法。该方法可以自动创建比仅基于进化的 GNP 和传统进化计算（遗传编程）表现出更好性能的程序，因为除了任务执行后执行的进化之外，该方法还可以在任务执行期间（在线学习）改进程序。