We propose Multi-Agent Reinforcement Learning with Partial Reinforcement Optimizatizer (MARL-PRO), a novel population-based optimizer that extends the Partial Reinforcement Optimizer (PRO) by incorporating cooperative multi-agent reinforcement learning. In MARL-PRO, each agent employs tabular Q-learning over a factorized discrete action space—subspace size, step scale, and direction (attraction or repulsion)—while a sensitivity vector guides updates toward the most responsive coordinates. A simple two-phase schedule further balances exploration and exploitation. Agents coordinate indirectly through shared incumbents and peer-based repulsion, fostering both convergence and diversity. Periodic local refinement and diversity resets enhance stability and prevent collapse. Evaluations on four benchmark functions (Sphere, Rosenbrock, Schwefel 1.2, and Schwefel 2.22) demonstrate that MARL-PRO consistently achieves faster and more stable convergence than PRO, particularly on ill-conditioned and nonseparable landscapes. These results highlight MARL-PRO’s potential as a scalable and interpretable framework for black-box optimization in distributed and cooperative environments.

Keywords