Reproduce “Beyond A*: Better Planning with Transformers via Search Dynamics Bootstrapping” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Lucas Lehnert, Sainbayar Sukhbaatar, DiJia Su, Qinqing Zheng, Paul Mcvay, Michael Rabbat, Yuandong Tian
This research from FAIR at Meta demonstrates training Transformers to learn the internal search dynamics of A* search, enabling them to solve complex planning tasks like maze navigation and Sokoban. Through a novel 'search dynamics bootstrapping' method, the models learn to generate optimal plans while executing significantly fewer internal search steps than the original A* algorithm, improving search efficiency.
Key insights
- Training Transformers on symbolic search dynamics, rather than just solutions, enables robust performance on complex multi-step planning tasks.
- The 'search dynamics bootstrapping' method allows Transformers to discover more efficient search strategies, generating optimal plans with fewer internal steps than the A* algorithm they imitate.
- Exposing the model to the 'thinking process' (search trace) of a planner leads to better generalization and data efficiency for Transformers in planning domains.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.