In response to the problem of premature convergence of AI (Artificial Intelligence) agent behavior caused by the decline of population diversity during the evolution process, this paper proposes a dynamic entropy constrained evolution strategy DGA-DECES. This method parameterizes the policy space through deep networks, integrates the global search of genetic algorithms with local optimization of reinforcement learning, and focuses on the dynamic entropy constraint mechanism: introducing entropy compensation term in fitness design, and maintaining gene diversity through adaptive selection pressure regulation. Experiments have shown that in non steady state multi-agent games and high-dimensional robot control tasks, the Shannon entropy of the population increases by 75%, and the convergence stability of the strategy is significantly better than that of the benchmark algorithm. Compared to NSGA-II, the optimization range of dynamic game tasks has increased by 30.2%; Compared to CMA-ES, the high-dimensional control error reduction rate has increased by 31.0%, verifying its efficiency and robustness in complex tasks.