Building Competitive Pokémon Agents: Learning, Search, and Adaptation Across Singles Formats
Year: 2026
Type: preprint
Abstract: An autonomous Pokémon singles agent reached rank 7 on the public Generation 9 UnderUsed ladder on 18 October 2026, with 1591.3 Elo and a recorded account score of 64 wins and 17 losses. Related agents recorded 1608.1 Elo at rank 49 in Ubers and 1664.6 Elo at rank 348 in OverUsed. The agents inferred hidden opponent sets from observed play and used learned move models to guide a bounded search over simulated battles. A separate Ubers evaluation recorded 138 wins in 192 completed games against a pinned Foul Play configuration using three seconds, one thread and search parallelism one (71.9%; reported 95% interval 65.1–77.8%). In Little Cup, replacing a scripted root prior with a human move model increased recorded wins from 371/426 to 420/450, although the schedule was incomplete. Conversely, pure mirror self-play and early learned evaluators regressed, and higher prediction accuracy did not consistently translate into stronger play. Adaptations across seven Generation 9 tiers and earlier-generation OU expose additional tradeoffs in model latency, hidden-team inference and rule fidelity. Ladder peaks are historical observations, development comparisons have stated confounders, and the available evidence does not establish superiority over Metamon or every configuration of Foul Play.
Author: Mamun, S M Abdulla Al
Institution: University of Manchester
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Keywords: autonomous agents, foul play, Uber
Open Access status: green
OpenAlex ID: W7236748410
