arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2402.14083cs.AI
收藏

超越 A*:通过搜索动态自举利用 Transformers 实现更优规划

Beyond A*: Better Planning with Transformers via Search Dynamics Bootstrapping

FAIR at Meta(Meta FAIR研究院)

Lucas Lehnert, Sainbayar Sukhbaatar, DiJia Su, Qinqing Zheng, Paul Mcvay, Michael Rabbat, Yuandong Tian

AI总结:

本研究提出通过训练 Transformer 模型预测 A* 算法的搜索动态,并微调得到 Searchformer,以更小的模型规模和数据量更高效地解决复杂规划任务。

摘要:

尽管 Transformers 在各种应用场景中取得了巨大进展,但在解决复杂决策任务时,此类架构仍落后于传统的符号规划器。在本研究中,我们展示了如何训练 Transformers 来解决复杂的规划任务。这是通过训练一个编码器-解码器 Transformer 模型来预测 $A^*$ 搜索算法的搜索动态而实现的。我们对该模型进行微调以获得 Searchformer,这是一个 Transformer 模型,在 93.7% 的情况下能最优地解决此前未见过的 Sokoban 谜题,同时比最初用于训练的 $A^*$ 实现少用高达 26.8% 的搜索步骤。在我们的训练方法中,$A^*$ 的搜索动态被表示为一个 token 序列,概述了在符号规划期间任务状态何时被添加到搜索树中以及何时从搜索树中移除。Searchformer 显著优于直接预测最优规划的基线模型,其模型尺寸小 5-10$\ imes$,训练数据集大 10$\ imes$。最后,我们展示了 Searchformer 如何扩展到更大、更复杂的决策任务,并提高了已解决任务的百分比以及缩短了搜索动态。

英文摘要:

While Transformers have enabled tremendous progress in various application settings, such architectures still trail behind traditional symbolic planners for solving complex decision making tasks. In this work, we demonstrate how to train Transformers to solve complex planning tasks. This is accomplished by training an encoder-decoder Transformer model to predict the search dynamics of the $A^*$ search algorithm. We fine tune this model to obtain a Searchformer, a Transformer model that optimally solves previously unseen Sokoban puzzles 93.7% of the time, while using up to 26.8% fewer search steps than the $A^*$ implementation that was used for training initially. In our training method, $A^*$'s search dynamics are expressed as a token sequence outlining when task states are added and removed into the search tree during symbolic planning. Searchformer significantly outperforms baselines that predict the optimal plan directly with a 5-10$\times$ smaller model size and a 10$\times$ smaller training dataset. Lastly, we demonstrate how Searchformer scales to larger and more complex decision making tasks with improved percentage of solved tasks and shortened search dynamics.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑