Boosting Perturbed Gradient Ascent for Last-Iterate Convergence in Games
Authors: Kenshi Abe, Mitsuki Sakamoto, Kaito Ariu, Atsushi Iwasaki
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we present the empirical results of our GABP, comparing its performance with APGA (Abe et al., 2024), OG (Daskalakis et al., 2018; Wei et al., 2021), and AOG (Cai & Zheng, 2023). We conduct experiments on two classes of concave-convex games. |
| Researcher Affiliation | Collaboration | Kenshi Abe1,2 Mitsuki Sakamoto1 Kaito Ariu1 Atsushi Iwasaki2 1Cyber Agent 2The University of Electro-Communications EMAIL EMAIL |
| Pseudocode | Yes | The pseudo-code of GABP is outlined in Algorithm 1. Algorithm 1 GABP for player i. |
| Open Source Code | Yes | An implementation of our method is available at https://github.com/Cyber Agent AILab/ boosting-perturbed-ga |
| Open Datasets | No | We conduct experiments on two classes of concave-convex games. One is random payoff games, which are two-player zero-sum normal-form games with payoff matrices of size d. [...] The other is a hard concave-convex game (Ouyang & Xu, 2021), formulated as the following max-min optimization problem: maxx X1 miny X2 f(x, y), where f(x, y) = 1 2x Hx + h x + Ax b, y . Following the setup in Cai & Zheng (2023), we choose X1 = X2 = [ 200, 200]d with d = 100. The precise terms of H Rd d, A Rd d, b Rd, and h Rd are provided in Appendix E.2. |
| Dataset Splits | No | For the random payoff games with full or noisy feedback, 50 payoff matrices are generated using different random seeds. Likewise, for the hard concave-convex games, we use 10 different random seeds. The paper describes the generation of problem instances but does not specify training/test/validation splits for a dataset. |
| Hardware Specification | Yes | The experiments were conducted on mac OS Sonoma 14.4.1 with Apple M2 Max and 32GB RAM. |
| Software Dependencies | No | The experiments were conducted on mac OS Sonoma 14.4.1 with Apple M2 Max and 32GB RAM. The paper only mentions the operating system but does not specify software libraries or frameworks with version numbers used for implementation. |
| Experiment Setup | Yes | The detailed hyperparameters of the algorithms, tuned for best performance, are shown in Table 2 in Appendix E.3. |