Equilibrium Planning and Payoff-Model Error: A Pokémon VGC Team-Preview Case Study
An equilibrium solver can solve the wrong game accurately. We examine this problem at team preview in Pokémon VGC, where each player chooses four members of a six-member roster and two leads. Fresh evaluations of 193 recorded previews assign identical values to different reserves behind the same leads. Archived payoff diagnostics are largely additive, allowing concentration on small differences in row strength. A historical bundled replacement using search payoffs, prior-regularised equilibrium and match history won 6 of 20 best-of-three sets against the original preview. A controlled follow-up completed four replicates comprising 320 sets. Leaf/RM, search/RM, leaf/regularised and search/regularised arms won 40, 43, 37 and 41 of 80 sets against a fixed leaf/RM reference. Clustered-bootstrap intervals for both marginal effects and the interaction include zero; the controlled comparison establishes neither improvement nor equivalence. In a perturbation analysis of recorded matrices, small noise increases concentration by breaking ties. Regularisation reduces added regret but has higher absolute regret on the unperturbed matrices at every tested scale. Repeated-call variability is much smaller than the historical cross-cell spread, weakening an earlier noise interpretation. The study separates three questions: whether the input represents the missing distinctions, whether the solver is stable, and what the games show. It introduces no new planning method. It does not show that regularisation is generally harmful.
Authors
- S M Abdulla Al Mamun
Institutions
- University of Manchester (GB)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24373819
- Primary Topic
- Artificial Intelligence in Games
- Type
- preprint