Extensive-Form Game Solving via Blackwell Approachability on Treeplexes
Authors: Darshan Chakrabarti, Julien Grand-Clément, Christian Kroer
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We provide an extensive set of experiments to compare our framework with several algorithmic benchmarks, including CFR+ and its predictive variant, and we highlight interesting connections between practical performance and the stepsize-dependence or stepsize-invariance properties of classical algorithms. |
| Researcher Affiliation | Academia | Darshan Chakrabarti IEOR Department Columbia University dc3595@columbia.edu Julien Grand-Clément ISOM Department HEC Paris grand-clement@hec.fr Christian Kroer IEOR Department Columbia University christian.kroer@columbia.edu |
| Pseudocode | Yes | Algorithm 1 Blackwell approachability on the treeplex, Algorithm 2 PTB+, Algorithm 3 Smooth PTB+ |
| Open Source Code | No | We plan to do so after the revision process. |
| Open Datasets | Yes | We conduct two sets of numerical experiments to investigate the performance of our algorithms for solving several two-player zero-sum EFG benchmark games: Kuhn poker, Leduc poker, Liar s Dice, Goofspiel and Battleship. Additional experimental detail is given in Appendix K. |
| Dataset Splits | No | The paper focuses on solving extensive-form games and evaluates convergence to Nash equilibrium, which does not typically involve explicit train/validation/test splits of game data in the way supervised learning models do. |
| Hardware Specification | No | The paper does not specify any hardware details (e.g., CPU, GPU models, memory) used for running the experiments. It only mentions general computational aspects like "computation time" in the context of CFR+ evaluation in Appendix E. |
| Software Dependencies | No | The paper does not provide specific software dependency details with version numbers (e.g., Python 3.x, PyTorch 1.x, or specific solver versions) that would be necessary for exact replication of the experimental environment. |
| Experiment Setup | Yes | For algorithms that are not stepsize invariant (Smooth PTB+ and SC-POMD), we try stepsizes in η {0.05, 0.1, 0.5, 1, 2, 5} and we present the performance with the best stepsize. For Smooth PTB+, we use R0 = 0.1. For both Ada Grad TB+ and Adam TB+, we use δ = 1 10 6, and for Adam TB+ we use β1 = 0.9 and β2 = 0.999. |