A multimodal multi-objective evolutionary algorithm assisted by long short term memory

When solving multimodal multi-objective optimization problems, the exploration ability and exploitation ability of algorithms are important for searching the equivalent Pareto optimal solution set (PSs). However, most of the traditional algorithms adopt meta-heuristic operators to reproduce offspring, which have superior exploration ability but insufficient exploitation ability. This leads to the incomplete and uneven distribution of the obtained PSs. To solve the above problem, an evolutionary algorithm based on Long Short Term Memory (LSTM) is proposed, which processes the population information under different generations into time series and learns the evolution regular patterns through LSTM to predict future population distribution. Specifically, an LSTM-based prediction model is constructed to reproduce promising offspring, which improves exploitation ability. Based on this, a reproduction strategy based on meta-heuristic operators and LSTM is utilized to ensure good exploration and exploitation in the decision space. Moreover, a convergence vector is designed to preserves both global PSs and local PSs by calculating the convergence relationship. The experimental results on 62 MMOPs show that the proposed algorithm performs better eight advanced MMOEAs, and its performance in the decision space and objective space is improved by 14.32% and 11.35% compared with the closest competitors. Finally, the algorithm is employed to the map-based test problem in engineering, showing superior performance.