Per-Domain Generalizing Policies: On Learning Efficient and Robust Q-Value Functions (Extended Version with Technical Appendix)
This work addresses the problem of efficient and robust policy learning for planning tasks, offering an incremental improvement over existing methods.
The paper tackles the challenge of learning per-domain generalizing policies by advocating for Q-value functions over state-value functions, which reduces evaluation cost by processing only the current state. It introduces regularization terms to improve Q-value learning, resulting in policies that outperform state-value policies across 10 domains and are competitive with the planner LAMA-first.
Learning per-domain generalizing policies is a key challenge in learning for planning. Standard approaches learn state-value functions represented as graph neural networks using supervised learning on optimal plans generated by a teacher planner. In this work, we advocate for learning Q-value functions instead. Such policies are drastically cheaper to evaluate for a given state, as they need to process only the current state rather than every successor. Surprisingly, vanilla supervised learning of Q-values performs poorly as it does not learn to distinguish between the actions taken and those not taken by the teacher. We address this by using regularization terms that enforce this distinction, resulting in Q-value policies that consistently outperform state-value policies across a range of 10 domains and are competitive with the planner LAMA-first.