Deep Reinforcement Learning for Adaptive Parameter Control in Differential Evolution for Multi-Objective Optimization

Evolutionary algorithms (EA) are efficient population-based stochastic algorithms for solving optimization problems. The performance of EAs largely depends on the configuration of values of parameters that control their search. Previous works studied how to configure EAs, though, there is a lack of a general approach to effectively tune EAs. To fill this gap, this paper presents a consistent, automated approach for tuning and controlling parameterized search of an EA. For this, we propose a deep reinforcement learning (DRL) based approach called ‘DRL-APC-DE’ for online controlling search parameter values for a multi-objective Differential Evolution algorithm. The proposed method is trained and evaluated on widely adopted multi-objective test problems. The experimental results show that the proposed approach performs competitively to a non-adaptive Differential Evolution algorithm, tuned by grid search on the same range of possible parameter values. Subsequently, the trained algorithms have been applied to unseen multi-objective problems for the adaptive control of parameters. Results show the successful ability of DRL-APC-DE to control parameters for solving these problems, which has the potential to significantly reduce the dependency on parameter tuning for the successful application of EAs.