emiliodavola/awale-alphazero-like

Awale AlphaZero-like

This model card documents an Awale policy/value network implemented in Julia with Flux.jl. The YAML metadata above comes from the release summary and should be treated as the source of truth for this bundle.

Release

• —Release ID: 20260730_130144
• —Commit SHA: bf679ecaf185b330d077a7eb84dade78dc5c8f20
• —Timestamp: 2026-07-30T15:42:46
• —Bundle kind: public_safe
• —Model export format: float32

Model Details

• —Architecture: mlp
• —Parameter count: 31559
• —Framework: Julia with Flux.jl
• —Input encoding: canonicalized board state as a 4x12 tensor (48 Float32 features)
• —Outputs: policy logits for 6 local actions and a scalar value in [-1, 1]
• —License: MIT
• —Release date: 2026-07-30T15:42:46

Usage

The model is a policy/value network trained by self-play. Load the bundled weights and run inference with Julia and Flux.jl:

using Awale
model = Awale.Model.load_public_model("artifacts/model_best.f32")
logits, value = Awale.predict_inference(model, Awale.initial_state())

predict_inference returns policy logits for the 6 local actions of the player to move and a scalar position value in [-1, 1].

Training Details

The network was trained with AlphaZero-style self-play: Monte Carlo Tree Search (PUCT) generates games, and the network is updated on sampled positions with policy and value targets. The training state reports last_iter = 550 iterations. Bundled training configuration: training = enabled.

Evaluation

Evaluation pits the trained network against a RandomAgent baseline with 400 MCTS simulations per move over 100 evaluation games. A checkpoint is promoted only when it reaches a decided win rate of at least 56% over 200 promotion games against the current best.

Metrics:

• —Best selection score: 61.7
• —Baseline win rate: 100
• —Final loss: 1.747
• —Selection current best rate: 53.06

Checkpoint status:

• —model_best: did not pass the promotion gate at the last selection
• —model_last: final run state; not subject to the promotion gate
• —model_final: final run state; not subject to the promotion gate

Limitations

• —The model was trained exclusively by self-play and has not seen human games.
• —Policy outputs cover 6 local actions; effective strength depends on the MCTS budget used at inference time.
• —Evaluation reflects the fixed RandomAgent baseline and the 400-simulation search budget.

Bundle contents

• —artifacts/model_best.f32: best checkpoint weights (promotion-gated)
• —artifacts/model_config.toml: model configuration snapshot
• —artifacts/model_final.f32: final checkpoint weights (final run state)
• —artifacts/model_last.f32: last checkpoint weights (final run state)
• —artifacts/training_config.toml: runtime configuration snapshot
• —artifacts/training_state.toml: training state snapshot
• —release_summary.toml: release metadata and evaluation metrics
• —manifest.toml: bundle manifest with integrity checksums
• —README.md: this model card

Code

• —Training scripts
• —Inference code
• —Evaluation scripts
• —Configuration files
• —Repository

Citation

If you use this model or repository in your work, please cite the repository:

@misc{awale2026,
  title = {Awale AlphaZero-like: self-play reinforcement learning for Awale},
  author = {Emilio Correa Dávola},
  year = {2026},
  howpublished = {\url{https://github.com/emiliodavola/awale}}
}

The model is released under the MIT license.