An efficient evolutionary algorithm based on deep reinforcement learning for large-scale sparse multiobjective optimization

Large-scale sparse multiobjective optimization problems (SMOPs) widely exist in academic research and engineering applications. The curse of dimensionality and the fact that most decision variables take zero values make optimization very difficult. Sparse features are common to many practical complex problems currently, and using sparse features as a breakthrough point can enable many large-scale complex problems to be solved. We propose an efficient evolutionary algorithm based on deep reinforcement learning to solve large-scale SMOPs. Deep reinforcement learning networks are used for mining sparse variables to reduce the problem dimensionality, which is a challenge for large-scale multiobjective optimization. Then the three-way decision concept is used to optimize decision variables. The emphasis is on optimizing deterministic nonzero variables and continuously mining uncertain decision variables. Experimental results on sparse benchmark problems and real-world application problems show that the proposed algorithm performs well on SMOPs while being highly efficient.