Studies on Adaptive and Learning Agents on Production System

为了在生产系统上实现具有自适应和学习功能的智能体系统，实际要求开发低水平智能体系统，而不是高水平智能体系统。在过去的两年里，我从高级智能的角度开发了具有自适应和学习功能的代理系统。然而，我所采用的一种学习方法需要大容量的内存和高计算速度。低水平智能的agent系统是指(1)agent关注其所在的本地环境，(2)agent与本地环境中的其他agent进行通信，(3)agent自己决定做什么，(4)整个系统达到其目的。为了开发这样的智能体系统，引入了自组织映射（SOM）的概念。SOM由神经元及其突触组成。它的神经元只通过更新局部突触来组织自己。当我们将神经元和突触分别视为生产智能体及其相关信息时，就有可能构建具有自适应低层次智能的智能体系统。将此概念应用于虚拟工厂中的车辆路径规划问题（VPM）和车辆交付规划问题（VDPP）。在VDP和VDPP上，车辆被视为agent。SOM请求agent之间的拓扑关系。在VDP中，采用直线拓扑。当给定智能体的起始位置和目标位置时，许多智能体随机分布在工厂中，它们之间具有直线拓扑关系。认识到SOM在起始点和目标点位置周围形成相交路径。然后，采用b样条插值方法中的多结点概念克服了这一缺陷。数值实验验证了agent之间只交换信息，最终生成了一个接近最优的车辆路径。对于VDPP， agent随机分布在工厂内，并在其中设置多个回路作为其拓扑结构。在每条环路上，多个代理被定位并固定在配送中心位置，就像一个多结。数值实验验证了智能体构建的近似等长环路作为车辆路径。这些结果发表在日本JSME和JSPE年会上。少

In order to realize an agent system with adaptive and learning functions on a production system, it is practically requested to develop the agent system with a low level intelligence rather than a high level intelligence. I have developed the agent system with adaptive and learning function from a viewpoint of the high level intelligence for the last two years. However, a learning method I have adopted needs a large capacity of memory and highly computing speed. The agent system with the low level intelligence means that (1) the agents concern his/her local circumstances, (2) they communicate with other agents within the local circumstances, (3) they make their decision what to do, and (4) the whole system attains its purpose. To develop such an agent system, a self-organizing maps (SOM) concept is introduced. SOM consists of neurons and their synapses. Its neurons proceed to organize themselves toward the arranged system by only updating their local synapses. When we regard the neuron … More s and synapses as production agents and their associated information, respectively, it becomes possible to construct the agent system with the adaptable low-level intelligence. This concept is applied to a vehicle path-planning problem (VPM) and a vehicle delivery-planning problem (VDPP) in a virtual factory.On VDP and VDPP, vehicles are regarded as the agents. SOM requests some topology relation between the agents. On VDP a straight-line topology is adopted for this purpose. When a start and goal locations are given to the agents, many agents are located in the factory at random and they have the straight-line topology relation. It is recognized that SOM makes intersection paths around the start and goal point locations. Then, multi-knot concept, which is used in the B-spline interpolation method, is employed to overcome this defect. Numerical experiments verify that the agents only exchange their information and finally generates a near optimum vehicle path. As for VDPP, the agents are located in the factory at random and multiple loops are set as their topology among them. Several agents are located and fixed at the delivery center location on each loop just as a multi-knot. Numerical experiments verify that the agents construct nearly equal-length loops as the vehicle paths. These results are presented in JSME and JSPE annual conference in Japan. Less