Under Construction: planned completion date October 2024

pkmn.ai

Baselines

No state eval beyond reducing HP of opponent, best matchup implicitly considers own HP vs opponents and multiple turns even though incredibly limited. Beyond is actual search

Agents

RandomPlayerRandomPlayer

TODO

RandomAttackRandomAttack

TODO

MaxDamagePlayerMaxDamagePlayer

TODO (favors HO) - needs a model (full in stats for opponent, use randbats heuristic of 85 evs or 252 evs depending on gen) - potentially need ability (Heatproof), item (AV)

MaxBasePowerMaxBasePower

TODO

MaxSimplifiedDamageMaxSimplifiedDamage

TODO

MaxEstimatedDamageMaxEstimatedDamage

TODO

BestMatchupPlayerBestMatchupPlayer

(favors Bulky Offense, can play HO) - switch if can take expected hit and KO faster than can be KOed - really needs some sort of opponent model - needs to extend model further than MDP to include potential moves

Enchancements

(compare benefit)


Run competition against each other Compare performance on randbats ladders (N=100?) and respective Elo/GXE - respectful laddering to not spam, N games over duration D window

Issues - long term planning - non damaging moves (boosting/hazards/recovery) - switching

@online{pkmn2024baselines,
  title  = "Competitive Pokémon Artificial Intelligence Concepts — Baselines",
  author = "pkmn.ai",
  month  = "March",
  year   = "2024",
  url    = "https://pkmn.ai/concepts/baselines/"
}