Inapproximability of Optimal Multi-Agent Pathfinding Problems

Authors: Xing Tan, Alban Grastien

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Theoretical This paper examines the computational approximability of optimal MAPF problems (i.e., minimizing makespan for agent travel distance and maximizing the total number of agents reaching their goals), providing a first set of several inapproximability results for these problems. The results reveal an inherent limitation in approximating optimal solutions for MAPFs, provide a deeper understanding regarding their computational intractability, thus offer foundational references for future research.
Researcher Affiliation Academia 1Department of Computer Science, Lakehead University, Canada 2Universit e Paris-Saclay, CEA, List, F-91120, Palaiseau, France EMAIL, EMAIL
Pseudocode No The paper describes reductions and proofs for theoretical complexity analysis but does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not contain any explicit statements about making source code available, nor does it provide links to code repositories.
Open Datasets No The paper focuses on theoretical problem instances (like 3DM and MAXE3SAT) used for reductions, not empirical datasets. No datasets are mentioned as being publicly available with access information.
Dataset Splits No The paper is theoretical and does not involve empirical experiments with datasets that would require training/test/validation splits.
Hardware Specification No The paper focuses on theoretical complexity analysis and does not describe any experimental setup that would involve specific hardware.
Software Dependencies No The paper is theoretical and does not describe any specific software or library dependencies with version numbers for implementation.
Experiment Setup No The paper is theoretical and does not present empirical experiments, therefore, there is no experimental setup, hyperparameters, or system-level training settings described.