A research on optimization of cellular neural networks based on evolutionary computation

我们提出了一种优化神经网络的方法--人工细胞神经网络(ACNN)。ACNN是一种神经网络，其中每个神经单元只与其相邻的单元连接。由于我们不需要连接任何相距较远的单元，所以我们可以很容易地制作二维和三维ACNN的LSI，在本研究中，神经元之间的连接权被限制在+1，-1和0三个值中的一个，并且ACNN中的每个连接权可以独立设置。因此，我们必须充分设置许多连接权，才能使ACNN对给定的任务起到很好的作用。我们使用遗传算法(GA)来进行优化。使用遗传算法，我们可以从具有随机设置的连接权的ACNN的初始种群中通过迭代生成用于各种信息处理任务的ACNN。提出了基于遗传算法的二维和三维人工神经网络的自动设计方法，通过定义适合于给定任务的评价函数，可以生成任意的人工神经网络。作为该方法的应用，我们处理了基于智能体的人工智能领域中的几个问题：首先，我们应用了2D ACNN来追踪问题。在这些问题中，虚拟追逐者的动作控制是由2D ACNN控制的，接下来，我们应用3D ACNN来解决迷宫问题。在这个问题中，三维ACNN的输入和输出分别是迷宫中主体的当前邻居和他的下一个移动方向。通过使用我们的方法，自动生成在最短时间内将代理从起始点移动到目标点的3D ACNN。我们开发了二维和三维ACNN的优化方法，并将其应用于人工智能。我们现在正计划制造2D和3D ACNN的LSI芯片。

We developed a method to optimize neural networks named "Artificial Cellular Neural Networks (ACNN)". ACNN is a kind of neural network in which each neural unit connects only with his neighboring units. Since we need not connect any units which are far away from each other, we can easily make LSIs of two-dimensional and three-dimensional ACNN, In this research, connection weight among neural units is restricted to one of the three values, +1, -1 and 0, and every connection weights in ACNNs can be set independently. Therefore, we have to set many connection weights adequately in order to make an ACNN work well for a given task. We employed Genetic Algorithm(GA) for this optimization. Using a GA, we can make ACNNs for a variety of information processing tasks through generation iterations from the initial population of ACNNs which have randomly set connection weights. We developed automatic designing methods for 2D and 3D ACNN based on GA.Using our method, we can generate any ACNNs b.y defining evaluation functions appropriate for given tasks. As applications of our method, we treated several problems in the field of agent based artificial intelligence, First, we applied 2D ACNN for chasing problems. In these problems, action control of a virtual chaser was controlled by a 2D ACNN, Next, we applied 3D ACNN for a maze problem. The input and output of a 3D ACNN in this problem are the current neighbor of an agent in a maze and his next moving direction, respectively. By using our method, a 3D ACNN which moves the agent from the start point to the goal point in the shortest time is automatically generated. We developed the optimization method for 2D and 3D ACNN and applied them to artificial intelligence. We are now planning to make LSI chips of 2D and 3D ACNN.