Subgoal Search

Subgoal search is a technique used to break down complex problems into smaller, more manageable sub-problems, improving efficiency and enabling the solution of long-horizon tasks. Current research focuses on developing methods to automatically generate relevant subgoals, often leveraging large language models or diffusion models to learn from limited expert data or even non-expert observations, and integrating these subgoals into various frameworks like hierarchical reinforcement learning and model predictive control. This approach is proving valuable across diverse fields, from robotics and theorem proving to solving complex mathematical problems, by improving learning efficiency and enabling the solution of tasks previously intractable for AI systems.