Fuzzy–Evolutionary Systems
Hugues Bersini · 2020
The crossover of evolutionary algorithms with fuzzy models has given birth to four offspring: first the use of fuzzy rules to make the behavior of evolutionary algorithms self-adaptive, second the use of evolutionary algorithms to improve the performance of fuzzy algorithms, third the addition of reinforcement learning to endow fuzzy sensory–motor autonomous agents with the capacity to learn on-line to receive more and more frequent reward, and finally the use of evolutionary algorithms to automatically tune the structure and parameters of fuzzy models. This chapter will essentially pay attention to the use of evolutionary algorithms for the automatic generation of fuzzy models. The factor which underlies the overparametrization of fuzzy models and consequently the need to automatically tune the structure and the parameters of these models is also, paradoxically, the same factor that makes this tuning not very hard to conceive. Fuzzy models are local models in the sense that each part of the model has a restricted and local responsibility, so that, although many of these parts are needed, one can easily detect where they are needed and how to distribute them. This fact also explains the combinatorial and multimodal nature of the search space of fuzzy models. Multiple combinations of fuzzy sets in input and output can provide solutions of similar quality. Any combinatorial optimization method which does not need too much prior information on the system to optimize and which works by stochastic and parallel exploration of the search space and by recombining or aggregating pieces of solutions is likely to occupy a high rank in the list of methods available for the structural tuning of fuzzy models. This explains the growing popularity gained nowadays by the use of evolutionary algorithms for the automatic data-based generation of fuzzy models.