Optimizing Collaborative Learning: A Hard-Constraint Reinforcement Learning Approach to Fair Team Composition

Abstract

Collaborative learning requires pedagogically balanced teams to function effectively. However, satisfying multiple strict constraints, such as specific gender ratios and minimized intra-group skill variance, transforms team composition into an NP-hard combinatorial optimization problem. Exact solving approaches, such as Optimization Modulo Theories (OMT), suffer from combinatorial explosion, taking hours to evaluate small cohorts (N = 40). We propose a Deep Reinforcement Learning (DRL) framework to resolve this bottleneck. Adapting the Long and Short-Term Constraints (LSTC) architecture, we enforce non-negotiable rules via dynamic action masking and cubic reward shaping. We utilize Deep Q-learning from Demonstrations (DQfD) via replay buffer pre-loading to initialize the policy with valid baselines. Empirical results show that our DRL agent achieves a significant computational speedup over the OMT-baseline while matching its engagement maximization. Furthermore, our approach eliminates the “failure clusters” prevalent in global optimization, improving worst-case team fairness. Robustness testing proves that the policy remains highly deterministic across randomized initializations.