GammaZero

GammaZero: Learning to Guide Belief-Space Search for Long-Horizon POMDPs (2025)

Accepted at ICAPS 2026

GammaZero is a graph neural network-based framework for partially observable Markov decision process (POMDP) planning that achieves dramatic computational efficiency gains while enabling zero-shot transfer to larger scenarios.

Key Contributions

• 20x Computational Efficiency: Reduced inference time from minutes to seconds while maintaining planning accuracy
• Zero-Shot Transfer: Enables generalization to scenarios 6x larger than those seen during training
• Uncertainty-Aware Graph Representation: Novel graph representation that captures partial observability and belief state uncertainty

Technical Approach

• Graph neural networks learn policy and value functions over belief-space representations
• Learned models integrated with Monte Carlo Tree Search (MCTS) to drastically reduce search requirements
• Uncertainty-aware node and edge features encode observation history and belief distributions

Applications

The framework is applicable to robotic planning, autonomous navigation, and other domains requiring decision-making under uncertainty with scalability to real-world problem sizes.