A Benchmark for Generalizing Across Diverse Team Strategies in Competitive Pokémon
Paper
• 2506.10326 • Published
This repo stores .zip checkpoints of runs from vgc-bench.
The training methods are described in VGC-Bench: Towards Mastering Diverse Team Strategies in Competitive Pokemon.
If the folder has name results<SEED>, it is just runs using that seed.
If the folder has name results<SEED>-<ID>, it contains a training run to learn a specific 2-team matchup and will have the teams stored inside.
The results1-san-antonio-2025-finals folder contains a behavior cloning-initialized self-play (BCSP) agent trained in the same way as the agent that was able to win a game against Aaron Traylor, a professional VGC competitor.
Cite us
@inproceedings{anglissvgc,
title={VGC-Bench: Towards Mastering Diverse Team Strategies in Competitive Pok{\'e}mon},
author={Angliss, Cameron L and Cui, Jiaxun and Hu, Jiaheng and Rahman, Arrasy and Stone, Peter},
booktitle={The 25th International Conference on Autonomous Agents and Multi-Agent Systems}
}