A Comparison of Multi-Objective Evolutionary Algorithms in Fuzzy Rule-Based Systems Generation

In this paper we compare three Pareto-based multi-objective evolutionary algorithms (MOEAs) plus a variant of one of them proposed by the authors. We use MOEAs to identify fuzzy rule-based systems (FRBS) of the Mamdani type from numerical data. We use two objective functions, namely accuracy and complexity, in order to find a good tradeoff between them. Each MOEA produces an approximation of the Pareto optimal front and the histogram of the number of generated solutions as a function of the complexity level. We evaluate their performance using two standard benchmarks in chaotic time series forecasting problems. Time consumption is also compared, distinguishing between the time spent in fitness computation and the time spent by each algorithm overhead. We show that the variant proposed by the authors gives the best Pareto front approximation, without adding significant overhead