A Large-Scale Multi-objective Evolutionary Algorithm Based on Two-layer Encoding Scheme for Neural Network Training

Evolutionary algorithms (EAs) have achieved great performance in solving multi-objective optimization problems (MOPs). However, there are more complex features in real-world MOPs. For example, the neural network training problem is characterized by large-scale decision variables and sparse Pareto-optimal solutions. This makes it difficult for most EAs to solve the neural network training problem. Therefore, this paper proposes a large-scale multi-objective evolutionary algorithm based on two-layer encoding scheme (MOEA- TES). In the proposed algorithm, the two-layer encoding scheme is employed. Moreover, a two-stage evolutionary strategy is designed for guiding the genetic operator. MOEA- TES is assessed by four real-world neural network training problems. Compared with five recent state-of-the-art algorithms, the proposed MOEA- TES achieves competitive results.