This project builds a Pokémon battle AI to investigate where the game's Nash equilibria
actually lie. There's no point keeping the findings to myself, so I'm sharing them here —
with playable, GPU-free demos of the learned policies.
Stage 3b
Cycling in an extreme matchup — Goodra-Hisui vs Cloyster
Goal. Build an extreme favorable/unfavorable matchup and see whether the decision
collapses to a single option — always switch, or always attack.
Setup. I paired a Special-Defense-focused Hisuian Goodra with a Defense-focused
Cloyster and tuned their movesets to create a perfectly favorable/unfavorable matchup.
The two share identical Attack and Special Attack, and every real stat is identical except
Defense and Special Defense, which are swapped between them. Shock Wave (Special / Electric /
60 BP) hits Cloyster for super-effective damage; Bulldoze (Physical / Ground / 60 BP) hits
Hisuian Goodra super-effectively — but each does very little to the other Pokémon, so
switching sharply cuts the damage taken. Movesets are mirrored across the two teams, so for
both sides "switching out of a bad matchup gives you a good one," making cycling battles
likely. Both Pokémon hold a Covert Cloak, which suppresses the added (secondary) effects of
moves. I trained the AI on this and studied the strategy it converged to.
Result. Even in a seemingly clear-cut matchup, surprisingly complex
mind games emerged. The tricky part: constrained by real moves, I couldn't build a perfectly
symmetric relationship — Bulldoze is neutral (1×) against Cloyster while Shock Wave is
resisted (½×) by Hisuian Goodra. As a result, "even in a favorable matchup, don't fire Shock
Wave — bait a switch and hit Bulldoze instead" became a very strong line.
Goal. Build a truly extreme favorable/unfavorable matchup and test whether the
decision collapses to a single option.
Setup. I paired a Special-Defense-focused Goodra with a Defense-focused Cloyster, then
gave them fictional moves — a 60-BP physical Fairy move and a 60-BP special Fighting move.
Real stats are identical and the type-effectiveness relationships are made perfectly
symmetric, to see what happens.
Result. It became considerably simpler, yet complex mind games still
emerged. Because you gain an edge by knocking out just one of the opponent's two Pokémon —
thereby denying them a switch — bait-switching to focus-fire a single target was effective.
In the end, a favorable matchup did not collapse to "always attack," nor an unfavorable one
to "always switch."
Neutral coverage in an extreme matchup — Goodra-Hisui vs Cloyster
Goal. Return to Stage 3b and give every Pokémon two neutral attacks — Crunch and
Dark Pulse — to see how a useful hit that does not exploit a weakness changes the equilibrium.
Setup. Species, stats, teams, and super-effective moves are identical to Stage 3b.
Every Pokémon now has its original coverage move plus 80-BP physical Crunch and 80-BP
special Dark Pulse.
Result. The equilibrium switch rate fell from 55.1% to 23.6%, and
average battles shortened from 30.0 to 11.0 turns. Yet strategy became more mixed: even in
favorable matchups, neutral attacks are used 54.7% of the time to punish an expected switch.
Neutral coverage in a truly extreme matchup — Goodra vs Cloyster
Goal. Stage 3d added neutral coverage to Stage 3b. Stage 3e applies the same change
to Stage 3c, completing a 2×2 grid. That grid isolates a question Stage 3d alone could not
answer: did 3d's dramatic collapse in switching come from the neutral attacks themselves, or
from the asymmetry that was baked into Stage 3b?
Setup. Species, stats, and the super-effective move assignment are identical to
Stage 3c. Every Pokémon now also carries 80-BP physical Crunch and 80-BP special Dark Pulse.
Because the matchup stays perfectly symmetric, the equilibrium must satisfy
V(s) + V(mirror of s) = 1 — a strict check on the solver that Stage 3d could not
provide. It holds to within one quantization step across all 3,380,000 states.
Result. The answer is the neutral attacks. Switching fell 54.1% →
24.5% and battles shortened 31.3 → 11.1 turns — almost exactly Stage 3d's numbers, with the
asymmetry removed. A neat surprise also appeared: an attack that looks strictly worse on
every damage roll is still played, because when the foe is nearly fainted, the excess damage
of the "better" move is wasted and only the damage carried into the switch-in matters.