Objective-Specific Multi-Agent Reinforcement Learning for Dynamic Multi-Objective Cooperative Optimization in Industrial Processes

Abstract

Dynamic industrial processes operate in time-varying environments with conflicting objectives, making realtime optimization difficult. Traditional scalarization methods use fixed objective weights and fail when conditions change. Evolutionary multi-objective algorithms can approximate the Pareto front but are too computationally expensive for real time. This paper proposes an objective-specific multi-agent reinforcement learning framework for dynamic multi-objective cooperative optimization. Each agent is assigned to one objective and learns its own policy, while an attention-based action fusion mechanism coordinates shared control decisions. This removes the need for manually chosen static reward weights and improves adaptive trade-offs under dynamic conditions. Experiments on real cement clinker calcination data with three objectives-minimizing unit heat consumption, minimizing unit electricity consumption, and maximizing clinker output-show that the proposed method outperforms the Weighted Sum, NSGA-II, and attention-based multi-agent baselines in Pareto solution quality, training stability, and convergence speed. The framework is powerful and scalable for real-world dynamic industrial multi-objective optimization.