When More Search Does Not Help: Search Effort and Playing Strength in Imperfect-Information Pokémon
Does more computation at decision time make a game-playing agent stronger? We study this question in a family of imperfect-information Pokémon search agents. The contribution is measurement, not a new search algorithm. A pre-specified budget experiment in Gen 4 OU crossed four decision budgets with four fixed model widths. It completed 6,148 of 6,176 requested games. The archived team-clustered estimate is +0.027 logit per doubling of iterations, with a 95% interval of [-0.007,+0.061]. An exploratory interpretation of the training-nonoverlap subset is more favorable to extra search. Two separate Ubers case studies illustrate a different issue. Native engine search recorded far more iterations of a different kind, yet its learned-leaf arm lost 7.0 percentage points relative to the reference search. A port preserving much of the reference search scored 3.8 points higher with a larger cap, but its trial was only 76.6% complete and its multiplicity-adjusted interval includes zero. An offline probe re-searched 400 real UU decisions at five budgets and three seeds. Agreement with the modal 6-second choice was 66.8% at 0.25 seconds and 88.3% at 3 seconds; the 6-second runs agreed with the modal choice in 88.7% of positions and with each other in 78.7% of seed pairs. This measures decision stability, not playing strength. These results distinguish budget scaling within a procedure from throughput comparisons across procedures. They do not show that adding iterations to a fixed algorithm makes it worse, nor identify a single cause of the native search's loss.
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.24373941
- Primary Topic
- Artificial Intelligence in Games
- Type
- preprint