pkmn/ai

The home of competitive Pokémon artificial intelligence

---

pkmn.ai is the home of competitive Pokémon artificial intelligence. The site serves as the definitive reference for existing competitive Pokémon AI projects and papers, relevant research, and explanations of important concepts and terms related to the intersection of competitive Pokémon battling and artificial intelligence while also providing authoritative rankings of existing agents which are continually tested within a controlled environment.

Researchers, developers, and interested onlookers are encouraged to join the pkmn.ai community on Discord, and all are welcome to request the inclusion of new agents on the leaderboard.

All of the code in this repository is distributed under the terms of the MIT License and all of the content is made available under CC BY-SA 4.0.

ai's Issues

Time

• https://www.chessprogramming.org/Time_Management
• how the Pokémon Showdown timer works (in particular, on ladder)
• tracking |inactive| + clock skew
• pondering
• anytime algorithms/iterative deepening

Abstraction

• data bucketing
  • Kalose HP buckets
  • Athena RNG bucketing
• quantization (eg. storing usage stats as "q8" 1/256 probabilities at rest)
• logical/contextual input logs?

Search

• tree bandit search + relation MCTS
• bandit algorithms
• (expecti)minimax
• alpha beta pruning
• search data structures
• parallelism

Evaluation

• https://www.chessprogramming.org/Evaluation
• analysis of existing HCE functions used by projects
• building blocks (hazards? # of remaining Pokémon, etc)
• evaluation through NN

Inference

• information sets + perspectives
• valid inferences based on protocol
• reverse damage calculation
• leveraging usage stats
  • usage after already accounting for lead]
• possibly enhancements to statistics (joint probabilities, "correlation matrices")
• combining inferences with heuristics

Regret

• CFR in Pokemon
• modifications to CFR (MC-CFR, CFR+) etc

Learning

• machine learning in Pokémon
• enhancements
• online learning (ie. during a game)
• NNUE in Pokémon
• embedding (poke2vec, pokemb)
• one shot learning on top of base model