Autonomous systems with explanatory internal models

具有内部模型的自治系统需要基于各种试验的学习算法、从训练结果中提取一般知识的方法、分层控制结构及其切换机制。为了满足这些要求，研究了以下内容。作为一种基于各种试验的学习工具，设计了以下方法：一种可以产生非确定性试验的混沌行为的控制机制，一种根据需要控制搜索范围的优化方法，一种找到具有理想特征的各种（次）最优解的遗传共生算法，以及一种调整其自身设计参数的优化程序。通过实例验证了它们的有效性。为了实现一般知识的提取，提出了带输入门的神经网络和径向基函数网络中节点合并的方法。这使我们能够以一种普遍适用的方式表示知识，这种方式具有较少数量的输入（条件）。这已经通过学习自主移动代理的行为得到了验证。还设计了带节点门的神经网络，实现了层次控制结构，同时实现了层次控制结构的切换机制。通过非线性系统控制实例验证了这一点。此外，在自主移动智能体的行为学习中，增加了一种机制，即提取适用于一般环境的规则。这使得智能体能够更快地适应新环境，这已经在实例中得到了证明。

Autonomous systems with internal models requires learning algorithms based on variety of trials, methods for extracting general knowledge out of trained results, and hierarchical control structure and its switching mechanisms. To meet these requisites, the following have been studied.As a tool for learning based on variety of trials, the following have been devised : a control mechanism of chaotic behaviors which can produce non-deterministic trials, an optimization method which controls the search range as required, genetic symbiosis algorithm that finds various(sub-)optimal solutions with desirable features, and an optimization procedure that tunes its own design parameters. Their validity has been examined by examples.To achieve extraction of general knowledge, the following have been proposed : neural networks with input gates and a method to merge nodes in Radial Basis Function networks. These enables us to represent the knowledge in a generally applicable way which has a fewer number of inputs(conditions). This has been verfied through learning of behaviors of autonomous moving agents.Neural networks with node gates have been also devised which can realize hierarchical control structure as well as its switching mechanism at the same time. This has been confirmed by examples of nonlinear system control. In addition, in learning of behaviors of autonomous moving agents, a mechanism has been added that extract rules applicable to general environments to are. This makes it possible for the agents to adapt new environment more quickly, which has been demonstrated in examples.