japlace: Modified Pokémon Showdown AI

A heavily modified version of japlace, using a patched poke-engine backend with a learned policy network and additional search improvements for Pokémon Showdown.

The project combines Laplace's existing battle AI with:

• Learned policy priors for MCTS
• Item, Ability, and Tera prediction
• Hidden-information posterior sampling
• Battle history features
• PUCT / regret-matching root search
• Optional opponent-action heuristics
• Policy training and self-play data generation
• Built-in A/B benchmarking

Setup

Build poke-engine

cd poke-engine-main/poke-engine/poke-engine-py
maturin develop --release --features poke-engine/gen9,poke-engine/terastallization

Enter the AI project

cd ../../Laplace-Pokemon-Showdown-AI

Activate your virtual environment:

.venv\Scripts\activate

Generate Training Data

python -m laplace.cli.gen_policy_data --battles 500 --workers 10 --det 4 --time-ms 60

Training data is written to:

data_policy/

Train the Policy Network

python -m laplace.cli.train_policy --epochs 30

The trained model is saved as:

models/policy_net.pt

Benchmark

Compare the trained policy against the baseline:

python -m laplace.cli.bench_ab --battles 100 --workers 10 --challenger-kwargs "{\"policy_model_path\":\"models/policy_net.pt\"}"

Play on Pokémon Showdown

Once models/policy_net.pt exists:

python -m laplace.cli.ladder --format gen9randombattle

The ladder client automatically loads the trained policy model.

Project Structure

japlace/
├── poke-engine-main/
│   └── poke-engine/
├── Laplace-Pokemon-Showdown-AI/
│   ├── laplace/
│   ├── models/
│   └── data_policy/
├── poke-engine-root-prior.patch
└── laplace-phase1-policy-net.patch

This repository is a custom japlace modification, combining the original Laplace battle system with new neural-policy, search, and hidden-information features.