A Training Data Recipe to Accelerate A* Search with Language Models

Recent works in AI planning have proposed to combine LLMs with iterative tree-search algorithms like A* and MCTS, where LLMs are typically used to calculate the heuristic, guiding the planner towards the goal. However, combining these techniques is not trivial : LM-based heuristics are quite weak, incurring a high computational cost without a significant performance improvement. Existing methods to learn these heuristics do not consider the requirements of the planner, and typically need a lot of compute. Thus, in this work, we propose a distribution to downsample training data by identifying relevant data points to learn a performant heuristic, while constraining computational costs. To arrive at this model, we disentangle the requirements of the planner, in our case A* search, from that of the language model to generalise on this task. Surprisingly, we find an overlap between their requirements; A* requires more accurate predictions on nodes near the goal, and LMs need the same set of nodes for effective generalisation. With these insights, we can quantify the contribution of each node towards accelerating A* search, and subsequently derive a training distribution for learning LM-based heuristics. Following a recent work, we conduct our experiments on two classical planning domains, maze navigation and sokoban, with two test splits per domain, and two conventional loss functions. We reduce the number of iterations required to find the solutions by upto 13x, with a wall-clock speed-up of upto 5x.