A conceptual framework for construction of autonomous systems and determination of specifications for neural network controller

We present a new framework for constructing autonomous systems, such as agent systems, and address an input variable selection for cart pole control utilizing the new framework. Definitions of three spaces (sensor-motor space, internal representation space, and evaluation function space) and discussions on the features of this framework are given in this paper. Based on the framework, we construct an input variable selection mechanism for a neural network (NN) controller for the cart pole control in the sensor-motor space. In the internal representation space, a decision tree is constructed by ID3, which is used for the input variable selection. The specifications of the NN controller and the learning mechanism of the NN controller, evolutionary programming (EP), are determined by using genetic algorithm (GA) incorporating the knowledge extracted by ID3. These three search algorithms are streamlined in the framework for easy tuning by the designer. Simulations are done to demonstrate an effective learning system with the input selection mechanism based on the framework