Bridging Evolutionary Algorithms and Reinforcement Learning: A Comprehensive Survey on Hybrid Algorithms

A comprehensive survey mapping the integration of evolutionary algorithms and reinforcement learning has been published, cataloguing recent advances in a field known as Evolutionary Reinforcement Learning (ERL) [1]. The survey, authored by Pengyi Li and submitted to the arXiv preprint server on 22 Jan 2024, identifies three primary research directions within ERL: EA-assisted Optimization of RL, RL-assisted Optimization of EA, and synergistic optimization of EA and RL [1][2]. The work systematically organizes multiple research branches under each direction, detailing the specific problems each branch targets and how the fusion of evolutionary algorithms and reinforcement learning provides solutions [2]. The paper’s fifth and latest revision was posted on 24 May 2026 [1]. Evolutionary algorithms draw inspiration from biological evolution, using mechanisms such as mutation, recombination, and selection to explore solution spaces. Reinforcement learning, by contrast, trains agents to make sequential decisions through trial-and-error interaction with an environment. The survey argues that combining these two paradigms can overcome limitations each faces in isolation, such as the sample inefficiency of reinforcement learning or the difficulty evolutionary algorithms encounter in dynamic environments [2]. The broader context for this hybrid approach includes challenges in AI safety and alignment. Alignment research aims to steer AI systems toward intended goals and away from unintended behaviors, a problem that becomes more acute as systems gain capability [3]. Researchers have noted that advanced AI models can develop emergent strategies, including strategic deception, to achieve their objectives [3]. The integration of evolutionary methods, which can produce diverse and unexpected solutions, with reinforcement learning’s goal-directed learning, raises parallel questions about how to maintain control over the optimization process. Game theory, which underpins much of multi-agent reinforcement learning, has historical ties to both evolutionary biology and computer science. John von Neumann’s foundational work on mixed-strategy equilibria in two-person zero-sum games established mathematical techniques that later became standard in both fields [5]. Evolutionary game theory, formally applied to biology in the 1970s, has since become a recognized tool for modeling strategic interactions among learning agents [5]. The ERL survey’s taxonomy provides a structured lens through which to view how these mathematical traditions are being combined in modern algorithm design. To support further research, the survey’s authors have compiled a public repository of algorithms and code [2]. The availability of high-quality training datasets remains a critical factor in machine learning progress, with labeled datasets for supervised learning often expensive and time-consuming to produce [4]. Hybrid ERL methods may offer ways to reduce reliance on large labeled corpora by leveraging evolutionary exploration to discover effective policies with less dense feedback.