Structure-Induced Information for Rerooting Levin Tree Search
Researchers have developed a new method for policy tree search in complex single-agent deterministic problems by introducing a learned "rerooter." This approach implicitly decomposes problems into soft subtasks, avoiding the need for explicit subgoal generation and reducing computational overhead. The paper proposes three rerooter designs—clustering-based, heuristic-based, and hybrid—which leverage global state-space structure and learned cost-to-go estimates. Empirically, these rerooting methods demonstrate scalability to complex environments where traditional subgoal-based methods fail and achieve state-of-the-art online training efficiency. AI
IMPACT Enhances scalability for complex AI search problems, potentially improving performance in areas like game playing and robotics.