Nash equilibrium estimation in competitive Pokémon using search and supervised learning
Keywords
Loading...
Authors
Issue Date
2022-07-08
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
The Pokémon main series role-playing video games revolve around catching and training different Pokémon to battle with. The competitive scene focusses solely on the game’s battling aspect in a player vs player setting, transforming it into a zero-sum, non-deterministic, simultaneous-move strategy game with imperfect information. We propose a method to estimate Nash equilibria strategies for competitive Pokémon battles. By combining search with an evaluation network, we can set up a payoff matrix for any given turn within a Pokémon battle to be used for Nash equilibrium calculation. Our evaluation network trained on a gen-8-ou dataset was able to correctly predict the outcome of a battle for randomly sampled states with an overall accuracy of 0.740. In battles against an open-source heuristic expectiminimax agent by Patrick Mariglia, our agent using the same heuristic evaluation achieved average win rates of 0.173 (control) with regular competitive teams and 0.618 when using simplified teams. Our agent using the trained evaluation network achieved an average win rate of 0.295 with the regular competitive teams and did not battle within the simplified setting. The results indicate that an increase in evaluation accuracy leads to better Nash equilibria estimation, with our current evaluation network being the bottleneck of this method. Future experiments are required to determine whether a sufficient level of evaluation accuracy for our method can be achieved.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen