[Paper Review] A Structured Prediction Approach for Generalization in Cooperative Multi-Agent Reinforcement Learning

Nicolas Carion, Gabriel Synnaeve+2 more|arXiv (Cornell University)|Oct 19, 2019

Reinforcement Learning in Robotics14 citations

TL;DR

This paper proposes a structured prediction approach for cooperative multi-agent reinforcement learning that uses a centralized optimization procedure with learned scoring models to assign agents to tasks, enabling zero-shot generalization to larger problem instances. The method outperforms strong baselines on StarCraft: Brood War scenarios with up to five times more agents and tasks than seen during training.

ABSTRACT

Effective coordination is crucial to solve multi-agent collaborative (MAC) problems. While centralized reinforcement learning methods can optimally solve small MAC instances, they do not scale to large problems and they fail to generalize to scenarios different from those seen during training. In this paper, we consider MAC problems with some intrinsic notion of locality (e.g., geographic proximity) such that interactions between agents and tasks are locally limited. By leveraging this property, we introduce a novel structured prediction approach to assign agents to tasks. At each step, the assignment is obtained by solving a centralized optimization problem (the inference procedure) whose objective function is parameterized by a learned scoring model. We propose different combinations of inference procedures and scoring models able to represent coordination patterns of increasing complexity. The resulting assignment policy can be efficiently learned on small problem instances and readily reused in problems with more agents and tasks (i.e., zero-shot generalization). We report experimental results on a toy search and rescue problem and on several target selection scenarios in StarCraft: Brood War, in which our model significantly outperforms strong rule-based baselines on instances with 5 times more agents and tasks than those seen during training.

Motivation & Objective

• To address the lack of generalization in centralized multi-agent reinforcement learning methods when scaling to larger problem instances.
• To exploit locality and coordination structure in multi-agent collaborative (MAC) problems to enable efficient learning and transfer.
• To develop a high-level agent-task assignment policy that generalizes across problem sizes without retraining.
• To model complex coordination patterns—such as spreading agents across distant tasks or grouping them for focused attacks—using structured prediction.
• To achieve real-time inference performance suitable for real-time strategy games like StarCraft: Brood War.

Proposed method

• The method decomposes the multi-agent policy into a high-level assignment policy and a low-level action policy, focusing on learning the former.
• Agent-task assignments are computed via a centralized optimization problem (inference procedure) whose objective is parameterized by a learned scoring model.
• The scoring model uses linear and quadratic terms: agent-task scores for immediate coordination and task-task scores for long-term coordination patterns.
• The inference procedure is polynomial-time, enabling efficient scaling to large numbers of agents and tasks.
• The model is trained on small MAC problem instances and then directly applied to larger, unseen instances (zero-shot generalization).
• Different combinations of inference procedures and scoring models are evaluated, including linear programming (LP), quadratic optimization (Quad), and AMax.

Experimental results

Research questions

• RQ1 Can a structured prediction approach with learned scoring models generalize effective multi-agent coordination policies from small to large problem instances?
• RQ2 How well can a centralized optimization procedure with learned scores capture complex coordination patterns such as focus-firing or spatial spreading?
• RQ3 Does the inclusion of quadratic terms in the objective function improve generalization in scenarios requiring long-term coordination?
• RQ4 Can the method achieve real-time performance in complex, real-time strategy games like StarCraft: Brood War?
• RQ5 How does the method compare to rule-based baselines in terms of zero-shot generalization across varying problem sizes?

Key findings

• The method achieved a median win rate of 0.94 in the m10v10 StarCraft scenario, significantly outperforming the best heuristic (0.83) and other baselines.
• In the w60v67 scenario, the Quad model achieved a win rate of 0.33, outperforming the best heuristic (0.13) and demonstrating effective zero-shot generalization.
• The quadratic inference procedure (Quad) significantly outperformed LP and AMax in scenarios requiring long-term coordination, such as the Wraith battle settings, where focus-firing is critical.
• In collision-prone scenarios like Zergling-Hydralisk (zh), the LP model performed well due to simpler coordination needs, while Quad generalized poorly, indicating sensitivity to environmental constraints.
• The algorithm ran in under 500ms per decision in the largest scenario (80 vs 82 units), enabling real-time gameplay in StarCraft: Brood War.
• The model generalized to instances with up to five times more agents and tasks than seen during training, demonstrating strong zero-shot generalization.