Meta-Reinforcement Learning for Cross-Domain Adaptation in Collaborative Decision Systems

Abstract

Generalizing collaborative strategies across heterogeneous tasks remains a major challenge for reinforcement-driven systems. This work examines a meta-reinforcement learning framework that enables rapid adaptation of coordination policies across domains. The model employs model-agnostic meta-learning (MAML) combined with PPO to learn a shared initialization that can be fine-tuned with limited task-specific data. Experiments are conducted on a composite dataset of 11,400 tasks spanning scheduling, routing, and resource optimization domains. The approach achieves a 23.1% improvement in zero-shot performance and reduces adaptation steps by 34.6% compared to task-specific training. Furthermore, cross-domain performance variance is reduced by 20.8%, indicating more stable transfer. These findings suggest that meta-learning provides an effective mechanism for scaling collaborative decision systems to diverse environments.