A fuzzy adaptive learning control network with on-line structure and parameter learning.

This paper addresses a general connectionist model, called Fuzzy Adaptive Learning Control Network (FALCON), for the realization of a fuzzy logic control system. An on-line supervised structure/parameter learning algorithm is proposed for constructing the FALCON dynamically. It combines the backpropagation learning scheme for parameter learning and the fuzzy ART algorithm for structure learning. The supervised learning algorithm has some important features. First of all, it partitions the input state space and output control space using irregular fuzzy hyperboxes according to the distribution of training data. In many existing fuzzy or neural fuzzy control systems, the input and output spaces are always partitioned into "grids". As the number of input/output variables increase, the number of partitioned grids will grow combinatorially. To avoid the problem of combinatorial growing of partitioned grids in some complex systems, the proposed learning algorithm partitions the input/output spaces in a flexible way based on the distribution of training data. Second, the proposed learning algorithm can create and train the FALCON in a highly autonomous way. In its initial form, there is no membership function, fuzzy partition, and fuzzy logic rule. They are created and begin to grow as the first training pattern arrives. The users thus need not give it any a priori knowledge or even any initial information on these. In some real-time applications, exact training data may be expensive or even impossible to obtain. To solve this problem, a Reinforcement Fuzzy Adaptive Learning Control Network (RFALCON) is further proposed. The proposed RFALCON is constructed by integrating two FALCONs, one FALCON as a critic network, and the other as an action network. By combining temporal difference techniques, stochastic exploration, and a proposed on-line supervised structure/parameter learning algorithm, a reinforcement structure/parameter learning algorithm is proposed, which can construct a RFALCON dynamically through a reward/penalty signal. The ball and beam balancing system is presented to illustrate the performance and applicability of the proposed models and learning algorithms.