Empirical Game Theoretic Analysis: A Survey
Authors: Michael P. Wellman, Karl Tuyls, Amy Greenwald
JAIR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We ran 10,000 simulations of all distinct profiles in this game, yielding the empirical game displayed in Table 1. The results are shown in Figure 11. |
| Researcher Affiliation | Collaboration | Michael P. Wellman EMAIL University of Michigan Ann Arbor, Michigan, USA; Karl Tuyls EMAIL Meta AI Paris, France; Amy Greenwald EMAIL Brown University Providence, Rhode Island, USA |
| Pseudocode | Yes | Algorithm 1: PSRO pseudocode [Lanctot et al., 2017]. |
| Open Source Code | No | The paper is a survey of Empirical Game Theoretic Analysis and does not claim to release source code for its own methodology or analysis. It discusses various algorithms and their implementations by other researchers but does not provide concrete access to its own source code. |
| Open Datasets | No | The paper describes simulated game environments (e.g., sequential auctions, Prisoners Dilemma) and references a "homogeneous-good model employed by Wellman et al. [2017]" for valuations, but does not provide concrete access information (links, DOIs, specific repositories, or formal citations for a dataset) for any publicly available or open dataset used in its own reported experiments or examples. It mentions "Alpha Zero Chess checkpoints" as an example, but without access details for that specific data. |
| Dataset Splits | No | The paper describes generating data through simulations (e.g., "We ran 10,000 simulations of all distinct profiles in this game") and varying the number of samples per profile for evaluation, rather than using predefined training, validation, or test splits from an existing dataset. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory specifications) used for running the simulations and analyses presented in its figures and tables. It mentions "seven CPUhours per sample" in the context of a cited work, not its own experimental setup. |
| Software Dependencies | No | The paper does not provide specific ancillary software details with version numbers (e.g., Python 3.x, PyTorch 1.x, specific solvers) used to conduct its own reported simulations and analyses. It is a survey discussing various tools and methods, but does not specify the software environment for its own work. |
| Experiment Setup | Yes | Consider a small version with n = g = 3, and three strategies, defined by shading factors: ρ {0.3, 0.5, 0.7}. Valuations are drawn as in the homogeneous-good model employed by Wellman et al. [2017], and the auction in each round is first-price. We ran 10,000 simulations of all distinct profiles in this game, yielding the empirical game displayed in Table 1. The game has three players and four goods, and three available bidding strategies: ρ {0, 0.5, 1}. Valuations are based on the homogeneous-good model of Wellman et al. [2017], with maximum values drawn from player-specific distributions. We generated 100 random games, and for each game evaluated the corresponding empirical game after various numbers of samples (100, 200, 500, and 1000) per profile. In our example, empirical game ˆΓ was estimated from m = 20 samples for each cell of the game matrix. |