Mastering Zero-Shot Interactions in Cooperative and Competitive Simultaneous Games

Authors: Yannik Mahlau, Frederik Schubert, Bodo Rosenhahn

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We perform an extensive evaluation of Albatross on a set of cooperative and competitive simultaneous perfect-information games. In contrast to Alpha Zero, Albatross is able to exploit weak agents in the competitive game of Battlesnake.
Researcher Affiliation Academia Yannik Mahlau 1 Frederik Schubert 1 Bodo Rosenhahn 1 1Department for Information Processing, Leibniz University Hannover, Germany.
Pseudocode Yes Algorithm 1 Training of Alpha Zero for simultaneous games
Open Source Code Yes To support reproducibility, all of our code as well as the trained models are open source1. 1https://github.com/ymahlau/albatross
Open Datasets Yes We evaluate Albatross in the Overcooked benchmark (Carroll et al., 2019) and the competitive game of Battlesnake (Chung et al., 2020).
Dataset Splits No The paper mentions 'training episodes' and evaluation on 'five different seeds' but does not provide explicit details about train/validation/test dataset splits with percentages or counts for reproducibility.
Hardware Specification Yes Unless specified otherwise, we used Nvidia RTX3090 GPU and 14 Intel Xeon Gold 6258R CPU for each GPU. Those numbers were chosen to optimally saturate the compute cluster used.
Software Dependencies No The paper mentions reimplementing Overcooked in C++ and discusses the use of Python (e.g., in the context of the environment), but it does not provide specific version numbers for any software dependencies or libraries.
Experiment Setup Yes To support reproducibility of our results, we report all hyperparameters used in our experiments. In this section, we list common hyperparameters used across all experiments. Hyperparameters, which differ between experiments are listed in Table 2.