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..
\(\varepsilon\)-Optimally Solving Two-Player Zero-Sum POSGs
Authors: Erwan escudie, Matthia Sabatelli, Olivier Buffet, Jilles Dibangoye
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We show empirically that point-based value iteration (PBVI) algorithms, applied via this reduction, produce ε-optimal strategies across a range of benchmark domains, consistently matching or outperforming existing state-of-the-art methods. |
| Researcher Affiliation | Academia | Erwan C. Escudie EMAIL University of Groningen Matthia Sabatelli EMAIL University of Groningen Olivier Buffet EMAIL Inria Nancy Grand-Est Jilles Steeve Dibangoye EMAIL University of Groningen |
| Pseudocode | Yes | Algorithm 1 PBVI for M (resp. M). Algorithm 2 Bounded Pruning. Algorithm 3 Redundant Informed Occupancy State Pruning. |
| Open Source Code | Yes | To foster reproducibility, the full codebase, including configuration files and experimental scripts, is available at Escudie et al. [2025]. URL https://git.lwp.rug.nl/e.c.escudie/Neur IPS-2025-zs-POSG. University of Groningen, version 1.0.0, accessed October 2025. |
| Open Datasets | Yes | We evaluate our method on a suite of established benchmarks for simultaneous-move partially observable stochastic games (POSGs): Adversarial Tiger, Competitive Tiger, Recycling, Mabc, Matching Pennies, and three Pursuit-Evasion variants. These benchmarks are among the most challenging in the POSG literature; see http://masplan.org/ for detailed descriptions. |
| Dataset Splits | No | The paper evaluates on benchmark problems that are game environments rather than traditional datasets, so explicit training/test/validation splits are not applicable or described. It mentions 'horizons ℓ {2, 3, 4, 5, 7, 10}' which refers to the planning horizon, not dataset splits. |
| Hardware Specification | No | The NeurIPS Paper Checklist states 'The paper provides the computer resources used to conduct experiments.' but the main body and appendices (E Experiments) do not specify concrete hardware details such as GPU/CPU models or memory. |
| Software Dependencies | No | The paper's NeurIPS Paper Checklist states 'While this is primarily a planning-focused paper, we nonetheless report the parameter settings used in our experiments for completeness and reproducibility.', but the main text or appendices do not specify software dependencies with version numbers. |
| Experiment Setup | Yes | For each benchmark, we compare three variants of our PBVI algorithm: PBVI1 (baseline, without pruning), PBVI2, and PBVI3 (both applying the bounded pruning scheme from Section 5). We benchmark against the HSVI implementation of Delage et al. [2023] and the CFR+ algorithm of Tammelin [2014]. Table 3 summarises results for the most computationally demanding horizons, reporting the final value reached by each algorithm and the exploitability of the resulting focal policy. Full results for all tested horizons ℓ {2, 3, 4, 5, 7, 10} are deferred to Table 3. |