This work proposes a hybrid multi-objective reinforcement learning framework for optimizing autonomous Pokémon battle agent strategies by combining Non-dominated Sorting Genetic Algorithm II (NSGA-II) with Deep Q-Learning (DQN). The approach addresses the challenge of balancing multiple conflicting objectives, including maximizing win probability, minimizing damage received, and improving overall battle efficiency. NSGA-II is employed to evolve and select Pareto-optimal strategy parameters, enabling effective exploration of the multi-objective search space, while Deep Q-Learning refines action policies through interaction with the battle environment. Experimental evaluations demonstrate that the proposed NSGA-II–DQN framework outperforms traditional single-objective and baseline reinforcement learning methods in terms of adaptability, strategic diversity, and multi-criteria performance. The results highlight the effectiveness of integrating evolutionary multi-objective optimization with deep reinforcement learning for complex sequential decision-making problems. This framework is extensible beyond game-based environments and can be applied to broader domains requiring efficient trade-offs among competing objectives.
Mohan Manoj Kumar Bonthu (Tue,) studied this question.