Nonlinear system control using self-evolving neural fuzzy inference networks with reinforcement evolutionary learning

• Authors: Cheng-Jian Lin;Cheng-Hung Chen
• Venue: Applied Soft Computing
• Year: 2011

This study presents a reinforcement evolutionary learning algorithm (REL) for the self-evolving neural fuzzy inference networks (SENFIN). By applying functional link neural networks (FLNN) as the consequent part of the fuzzy rules, the proposed SENFIN model combines orthogonal polynomials and linearly independent functions in a functional expansion of the FLNN. The SENFIN model can generate the consequent part of a nonlinear combination of the input variables. An efficient reinforcement evolutionary learning algorithm (REL), which consists of structure learning and parameter learning, is also presented. The structure learning is to determine the number of fuzzy rules. It adopts a subgroup symbiotic evolution to yield several variable fuzzy systems and uses an elite-based structure strategy to find the suitable number of fuzzy rules for solving a specific problem. The parameter learning is to adjust parameters of the SENFIN. It is a hybrid evolutionary algorithm, i.e., combining the cooperative particle swarm optimization and the cultural algorithm, called the cultural cooperative particle swarm optimization (CCPSO). As the result, the CCPSO approach can increase the global search capacity by using the belief space. In this paper the proposed NFIN with an efficient reinforcement evolutionary learning algorithm had been evaluated by two reinforcement learning applications, i.e., to balance the cart-pole system and the ball and beam system. Experimental results have demonstrated that the proposed approach performs well in reinforcement learning problems.