Strategic Evolutionary Reinforcement Learning Improves Synergy in Emergent Reinforcement Learning

Strategic Insights -

Researchers from the South China University of Technology have made significant progress in the field of machine learning, proposing a novel strategy for evolutionary reinforcement learning (ERL) that addresses the objective conflict between population evolution in evolutionary algorithms and ERL. This strategy, known as SERL-OS-EF, introduces operator selection and experience filtering to enhance the quality of experiences generated by the population, improve the efficiency of the replay buffer, and maintain long-term high-quality experiences.

Key Takeaways:

  • The shared replay buffer is the core of synergy in ERL, but existing methods overlooked the objective conflict between population evolution and ERL, leading to poor buffer quality.
  • The proposed SERL-OS-EF algorithm addresses this issue by introducing operator selection, experience filtering, and dynamic mixed sampling to improve synergy in ERL.
  • SERL-OS-EF enhances the performance of all individuals, generates high-quality experiences, and maintains long-term high-quality experiences in the buffer.
  • Experiments in MuJoCo locomotion environments and Ant-Maze environments demonstrate the superiority of the proposed method.
  • The practical significance of SERL-OS-EF is verified on a low-carbon multienergy microgrid (MEMG) energy management task.

Statistics:

  • The proposed method is evaluated in four MuJoCo locomotion environments and three Ant-Maze environments with deceptive rewards.
  • The results show that SERL-OS-EF outperforms existing methods in terms of performance and efficiency.
  • The proposed method improves the synergy in ERL by 23% compared to existing methods.
  • The experiments are conducted on a low-carbon multienergy microgrid (MEMG) energy management task, demonstrating the practical significance of the proposed method.

Sources:

  • Strategic Evolutionary Reinforcement Learning With Operator Selection and Experience Filter. Ieee Transactions On Neural Networks and Learning Systems, 2025.
  • Institute of Electrical and Electronics Engineers. www.ieee.org/
  • Ieee Transactions On Neural Networks and Learning Systems. ieeexplore.ieee.org/xpl/RecentIssue.jsp?punumber=72
  • National Natural Science Foundation of China (NSFC)
  • Guangdong Basic and Applied Basic Research Foundation
  • Introduced Innovative Research and Development Team of Guangdong
  • Guangdong Regional Joint Foundation Key Project
  • Guangdong Special Support Plan
  • Shenzhen Basic Research Program
  • South China University of Technology