A Hybrid Optimization Framework for Dynamic Multi-Objective Problems Using NSGA-II and Deep Reinforcement Learning

Dynamic multi-objective optimization problems (DMOPPs) are prevalent in complex engineering and management systems, where decision-makers must balance conflicting objectives under dynamically changing conditions. Existing methods often struggle to simultaneously achieve global search capabilities and real-time adaptability. To address this challenge, this paper proposes a novel hybrid optimization framework that synergizes the strengths of evolutionary algorithms and deep reinforcement learning. The framework is structured around three core components: a multi-objective optimization model to define the problem space, a multi-agent game-theoretic model to capture strategic interactions between system entities, and a spatial-aware resource allocation module. We employ the Non-dominated Sorting Genetic Algorithm II (NSGA-II) to explore the global Pareto-optimal solution space in an offline stage. Subsequently, a Deep Deterministic Policy Gradient (DDPG) agent, a form of deep reinforcement learning, is utilized for online, fine-grained policy adjustments in response to real-time system feedback. The framework's efficacy is validated on a complex system simulation, demonstrating its ability to effectively manage system trade-offs, mitigate negative externalities (e.g., congestion or resource depletion), and maintain high operational efficiency. This work provides a robust and generalizable paradigm for solving a significant class of DMOPPs.