Local and Adaptive Mirror Descents in Extensive-Form Games
Authors: Côme Fiegel, Pierre Ménard, Tadashi Kozuno, Remi Munos, Vianney Perchet, Michal Valko
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We implemented Local OMD, with the parameters of Theorem 4.1 and Theorem 4.2, then tested it against the theoretically optimal Balanced CFR (Bai et al., 2022) using the balanced policy as the sample policy, and Balanced FTRL (Fiegel et al., 2023). The algorithms were compared on three standard benchmark games: Kuhn poker (Kuhn, 1950), Leduc poker (Southey et al., 2005) and liars dice, using the version 1.4 of the Open Spiel library (Lanctot et al., 2019) under the Apache 2.0 license. |
| Researcher Affiliation | Collaboration | Côme Fiegel CREST Fair Play, ENSAE Paris Palaiseau, France come.fiege@normalesup.org Pierre Ménard ENS Lyon Lyon, France Tadashi Kozuno OMRON SINIC X Tokyo, Japan Rémi Munos Google Deep Mind Paris, France Vianney Perchet CREST Fair Play, ENSAE Paris, Criteo AI Lab Paris, France Michal Valko INRIA |
| Pseudocode | Yes | Algorithm 1 Learning procedures with fixed sampling policies for two players; Algorithm 2 Local OMD |
| Open Source Code | Yes | The code is available at https: //github.com/anon5493/Local OMD-experiments. |
| Open Datasets | Yes | The algorithms were compared on three standard benchmark games: Kuhn poker (Kuhn, 1950), Leduc poker (Southey et al., 2005) and liars dice, using the version 1.4 of the Open Spiel library (Lanctot et al., 2019) under the Apache 2.0 license. |
| Dataset Splits | No | The paper does not specify explicit training, validation, or test dataset splits. It mentions learning over a 'total number of episodes' and 'self-play framework' for policy updates, which differs from traditional dataset splitting for supervised learning. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., CPU, GPU models, memory, or cloud resources) used for running the experiments. |
| Software Dependencies | Yes | using the version 1.4 of the Open Spiel library (Lanctot et al., 2019) under the Apache 2.0 license. |
| Experiment Setup | Yes | The learning rates (and the IX parameters for the relevant algorithms) were optimized independently for each algorithm using a grid search. |