Rapid Learning in Constrained Minimax Games with Negative Momentum
Authors: Zijian Fang, Zongkai Liu, Chao Yu, Chaohao Hu
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on both Normal Form Games (NFGs) and Extensive Form Games (EFGs) demonstrate that our momentum techniques can significantly improve algorithm performance, surpassing both their original versions and the SOTA baselines by a large margin. We conduct comprehensive experiments over randomly generated NFGs and four standard EFGs, including Kuhn Poker, Leduc Poker, Goofspiel and Liar s dice. The experimental results demonstrate that the momentum-augmented algorithms exhibit significant improvements over both their original versions and other existing strong last-iterate convergent baselines. |
| Researcher Affiliation | Academia | 1School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China 2Pengcheng Laboratory, Shenzhen, China 3Shanghai Innovation Institute, Shanghai, China EMAIL, EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1: Restarting Aggregated Momentum (RAM) ... Algorithm 2: Momentum RM+ (Mo RM+) |
| Open Source Code | Yes | Code https://github.com/kkkaiaiai/NM-Method |
| Open Datasets | Yes | We conduct comprehensive experiments over randomly generated NFGs and four standard EFGs, including Kuhn Poker, Leduc Poker, Goofspiel and Liar s dice. In the tabular setting of EFGs, we utilize games implemented in Open Spiel (Lanctot et al. 2020) |
| Dataset Splits | No | The paper describes using randomly generated NFGs and standard EFGs (Kuhn Poker, Leduc Poker, Goofspiel, Liar's dice) implemented in Open Spiel, but does not specify any training/test/validation splits for these games or generated data. |
| Hardware Specification | No | The paper does not contain specific details about the hardware (e.g., GPU/CPU models, memory) used to run the experiments. |
| Software Dependencies | No | The paper does not provide specific version numbers for any software dependencies (e.g., programming languages, libraries, frameworks, solvers) used in the experiments. |
| Experiment Setup | Yes | Detailed game description and hyper-parameters can be found in the Appendix B. The learning rate η is fixed at 2 in this context. For RM+ and Mo RM+, we employ alternating updates, while other algorithms are kept with their default settings, i.e., simultaneous updates. We set the uniform strategy as the initial magnet strategy for MMD-M. |