GNN-Based Action Ranking

GNN-Based Action Ranking for Classical Planning (2023)

This project develops a learning-to-rank approach for guiding classical planners using graph neural networks, achieving strong generalization with minimal training data.

Published at NeurIPS 2025

Key Results

• 89-100% success rates with strong out-of-distribution generalization
• 75% less training data required compared to value-function methods
• 13x better coverage on challenging problems vs. learning value-based methods
• Generalizes to test problems 8x larger than training problems
• Outperforms LLMs (GPT-O3, Gemini-2.5-Pro) by 20-30x in coverage and 4-5x in plan quality

Technical Approach

• Action-centric graph representation capturing planning state structure
• Graph neural network combined with autoregressive GRU decoder for action sequence prediction
• Learning to rank formulation rather than value estimation

Why Action Ranking?

Traditional approaches learn value functions or policies, which require extensive training data. Our ranking approach focuses on relative action quality, enabling efficient learning from smaller datasets while achieving better generalization.