Evolutionary Dual-Guided Deep Reinforcement Learning
Zikai Zhao, Xuesong Gao, Qiuxuan Wu, Yuxin Chen, Sancheng Li, Zhiyuan Han, Ruyu Zhang · 2025
In the field of evolutionary reinforcement learning algorithms, the combination of evolutionary algorithms and reinforcement learning includes direct policy search methods and indirect experience-guided methods. Direct policy search methods: Evolutionary algorithms directly optimize policy parameters. They have simple computations but require complex parameter tuning, as different tasks may need different search strategies. Indirect experience-guided methods: Evolutionary algorithms indirectly and singularly guide deep reinforcement learning algorithms through an experience pool. Although they can optimize an agent's policy to some extent, they have notable issues such as slow convergence speed, delayed information transmission, and difficulty in effectively handling high-dimensional state and action spaces. These problems mainly arise from the single guidance signal or indirect nature of the evolutionary process, leading to inefficient and unstable optimization. To address these issues, this paper proposes a Dual-Guidance Deep Evolutionary Reinforcement Learning Algorithm (DDERL).