自律システム構成法の枠組に関する一提案と入力情報選択による倒立振子制御

概要

論文の詳細を見る
We present a framework for constructing autonomous system, such as agent system and mobile robots and address a learning problem of input selection for cartpole control using neural network. The framework consists of three perspectives for the generalized functions of system: sensor-actuator space, internal representation space, and evaluation function space. Based on the framework, we construct an input selection mechanism for neural network learning by using Evolutionary Programming (EP) for the cartpole control. The specifications of the neural network controller are determined by using Genetic Algorithm (GA), which specifies the input signals, the number of hidden units, the control cycles, and the number of learning epochs for EP. Implementing the input selection by GA, we use the root attribute which is derived from decision tree given by ID3 from data in the sensor-actuator space. Simulations are done to demonstrate an effective learning system with the input selection mechanisms based on the framework.
社団法人日本機械学会の論文
1999-11-25