Abstract
This paper addresses the problem of self-optimization in intelligent agents operating in complex and dynamic environments and proposes an evolutionary strategy-driven adaptive learning algorithm. Centered on evolutionary dynamics, the method integrates fitness normalization, mutation constraints, and adaptive regularization mechanisms to construct a unified optimization framework with global search and self-regulation capabilities. In the algorithm design, multi-level fitness modeling and structured evolutionary operators are introduced, enabling the agent to achieve stable policy evolution and dynamic convergence under non-stationary distributions. The study first analyzes key factors influencing diversity preservation, selection balance, and convergence stability in evolutionary strategies, followed by sensitivity experiments under various data perturbation and anomaly ratio conditions. The results show that the proposed algorithm demonstrates strong robustness and generalization under reward noise, abnormal trajectories, and environmental uncertainty. The fitness normalization strategy effectively mitigates selection bias among individuals and enhances the stability of policy evolution, while the regularization mechanism further improves the balance and efficiency of global search. Overall, the findings confirm that the evolutionary strategy-driven self- optimization framework can achieve multidimensional stable evolution in complex systems, providing a transferable theoretical and methodological foundation for decision optimization and autonomous learning in intelligent agents.
