One Search Agent, Seven Ladders: Peak and Sustained Performance in Competitive Pokémon Singles
Year: 2026
Type: preprint
Abstract: We describe a search agent for competitive Pokémon singles and report how it performed on the public Pokémon Showdown ladders. The agent keeps a belief over the opponent's hidden sets, rebuilds sampled complete states in a fast simulator, and runs a time-budgeted tree search over public histories, guided by models of human play learned from replays. One method was adapted to seven Generation 9 singles tiers by changing the team, the candidate-set bank, the training replays and the time budget, within a shared search code base. Its per-tier versions entered the top 500 of all seven of these ladders, with best recorded ranks of 7 (UU), 24 (LC), 49 (Ubers), 77 (NU), 119 (RU), 222 (PU) and 348 (OU). Because peaks are selected maxima, we also report every autonomous game in the run logs (3,326 games, 65.7% won): the level the agent sustained, measured as a performance rating against the opponents it met, is above the tier's top-500 threshold in three tiers and below it in four, and a secondary simulation based on the recorded ratings places every logged peak within its central range. In side-swapped local matches on 250 team pairs the agent showed no detectable difference from Metamon's Kakuna model (244–256) but was weaker than the search agent Foul Play: 44.7% at Foul Play's own setting and 43.1% at a matched decision time.
Author: Mamun, S M Abdulla Al
Institution: University of Manchester
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236849127
