Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Decentralized Monte Carlo Tree Search for Partially Observable Multi-Agent Pathfinding
Authors: Alexey Skrynnik, Anton Andreychuk, Konstantin Yakovlev, Aleksandr Panov
AAAI 2024 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The experimental results show that our approach outperforms state-of-the-art learnable MAPF solvers. Empirical Evaluation Experimental Setup To evaluate the efficiency of MATS-LP, we have conducted a set of experiments, comparing it with existing learnable approaches tailored to solve LMAPF problems. |
| Researcher Affiliation | Academia | Alexey Skrynnik1,2, Anton Andreychuk1, Konstantin Yakovlev2,1, Aleksandr Panov1,2 1AIRI, Moscow, Russia 2 Federal Research Center for Computer Science and Control of Russian Academy of Sciences, Moscow, Russia |
| Pseudocode | No | The paper describes the steps of the MCTS algorithm (Selection, Expansion, Backpropagation) in narrative text without presenting them in structured pseudocode or an algorithm block. |
| Open Source Code | Yes | The source code is available at https://github.com/AIRI-Institute/mats-lp. |
| Open Datasets | No | The paper uses custom-generated maps and instances, and while it references a generator from another work ([Damani et al. 2021]), it does not provide concrete access information (link, DOI, or formal citation for the *specific dataset instances* used) to make these datasets publicly available or open. |
| Dataset Splits | No | The paper describes the experimental setup and training parameters but does not explicitly provide information on training, validation, and test dataset splits (e.g., percentages, sample counts, or predefined split references). |
| Hardware Specification | Yes | In total, we conducted 100 algorithm runs, which roughly corresponds to 120 GPU hours using a single Titan RTX. |
| Software Dependencies | No | The paper mentions using an 'open-sourced asynchronous implementation of the PPO algorithm' (referencing a GitHub repository), but it does not specify version numbers for this or any other key software dependencies (e.g., Python, PyTorch, etc.). |
| Experiment Setup | Yes | The episode length was set to 512 in all the experiments. All the agents had the same parameters: their field-of-view was 11 11, all possible actions were considered only for the closest 3 agents, including the main agent, γ-value was set to 0.96, the number of expansions per iteration 250, coefficient c was set to 4.4. More details and the values of the rest parameters are given in the Hyperparameters section below. Table 1 presents the hyperparameters of COSTTRACER and MATS-LP approaches. |