TransPath: Learning Heuristics for Grid-Based Pathfinding via Transformers
Authors: Daniil Kirilenko, Anton Andreychuk, Aleksandr Panov, Konstantin Yakovlev
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct a thorough empirical evaluation on a comprehensive dataset of planning tasks, showing that the suggested techniques i) reduce the computational effort of the A* up to a factor of 4x while producing the solutions, whose costs exceed those of the optimal solutions by less than 0.3% on average; ii) outperform the competitors, which include the conventional techniques from the heuristic search, i.e. weighted A*, as well as the state-of-the-art learnable planners. |
| Researcher Affiliation | Academia | Daniil Kirilenko1, Anton Andreychuk2, Aleksandr Panov1, 2, Konstantin Yakovlev1, 2 1 Federal Research Center for Computer Science and Control of Russian Academy of Sciences, Moscow, Russia 2 AIRI, Moscow, Russia anedanman@gmail.com, andreychuk@airi.net, panov@airi.net, yakovlev@isa.ru |
| Pseudocode | Yes | Fig. 2 shows the pseudocode of a generic heuristic search algorithm. |
| Open Source Code | Yes | The source code of our planners is publicly available at https: //www.github.com/AIRI-Institute/Trans Path |
| Open Datasets | Yes | We have adopted the TMP (Tiled Motion Planning) dataset that was used in (Yonetani et al. 2021) for empirical evaluation. This dataset is a modification of the MP dataset used in (Bhardwaj, Choudhury, and Scherer 2017). |
| Dataset Splits | Yes | They were divided in the proportion of 8:1:1 for train, validation, and test subsets in such a way that all augmented versions of the same map were presented only in one of the subsets. |
| Hardware Specification | Yes | It took us 3.5 hours to train each model on NVIDIA A100 80GB GPU. The prediction step for the batch size of 64 and the native torch float32 type required 9.5ms on Tesla A100 GPU (and 40ms on GTX 1660S). |
| Software Dependencies | No | The paper mentions 'Python', 'Adam optimizer', 'One Cycle LR learning rate scheduler', and 'torch float32' but does not specify version numbers for any software libraries or dependencies. |
| Experiment Setup | Yes | We train each model using Adam optimizer (Kingma and Ba 2014) for 35 epochs with a batch size of 512 and One Cycle LR learning rate scheduler (Smith and Topin 2018) at a maximum learning rate of 4 · 10−4. We use L2 loss for cf-values, pp-values and L1 loss for the cost-to-go estimates following (Takahashi et al. 2019). |