Automatic construction of feedforward/recurrent fuzzy systems by clustering-aided simplex particle swarm optimization

This paper proposes a new approach for automating the structure and parameter learning of fuzzy systems by clustering-aided simplex particle swarm optimization, called CSPSO. Unlike most evolutionary fuzzy systems, where the structure of the fuzzy system is assigned in advance, an on-line fuzzy clustering approach is proposed for system structure learning. This structure learning not only helps determine the number of rules automatically, but also avoids the generation of highly similar fuzzy sets on each input variable. In addition, it improves subsequent parameter learning performance by assigning suitable initial locations of the fuzzy sets on each input variable. Once a new rule is generated, the corresponding parameters are further tuned by the hybrid of the simplex method and particle swarm optimization (PSO). In CSPSO, each fuzzy system corresponds to a particle in PSO, and the idea of the simplex method is incorporated to improve PSO searching performance. To verify the performance of CSPSO, two simulations on feedforward fuzzy systems design are performed. In addition, design of a recurrent fuzzy controller for a practical experiment on water bath temperature control is performed. Comparisons with other design approaches are also made in these examples.