Bluefir/escampe-dataset

Escampe Game State Evaluation Dataset

Dataset Summary

This dataset contains millions of evaluated game states for the board game Escampe. It is designed to train Deep Neural Networks (such as the BandDPER architecture) to accurately predict the heuristic value of any given board state.

Escampe is a deterministic, perfect-information 6x6 board game played with Unicorns and Paladins. The dataset was procedurally generated using an Alpha-Beta search engine in Java that plays games of Escampe against itself, taking random opening positions and semi-random moves to guarantee high diversity in board states. Every position in this dataset has been deeply evaluated by the minimax engine using various heuristic configurations (Default, SPSA, Bayes, and Simplex) to assign a continuous score.

Dataset Structure

Data Fields

Each JSON entry corresponds to a specific board state from a game in progress. The fields include:

• white_paladins (int): A 64-bit integer bitboard mask representing the positions of White's Paladins.
• black_paladins (int): A 64-bit integer bitboard mask representing the positions of Black's Paladins.
• white_unicorn (int): The 0-35 index representing the square of White's Unicorn.
• black_unicorn (int): The 0-35 index representing the square of Black's Unicorn.
• required_band (int): An integer (1, 2, or 3) representing the required liseret/band the current player must start their move on. A 0 means no band constraint (free move).
• white_to_move (bool): true if it is White's turn to move, false otherwise.
• score (float): The target evaluation score normalized to the range [-1.0, 1.0]. The score is relative to the current player to move. A score of +1.0 indicates a guaranteed win (mate), while -1.0 indicates a guaranteed loss.

Data Instances

{
  "white_paladins": 8913048,
  "black_paladins": 40298872832,
  "white_unicorn": 14,
  "black_unicorn": 34,
  "required_band": 2,
  "white_to_move": true,
  "score": -1.0
}

Dataset Creation

The dataset was generated using a high-performance, multi-threaded Java minimax engine utilizing Alpha-Beta pruning.

• Diversity: Random openings are selected from a predefined pool of Escampe openings. Games proceed using a mix of best-moves and random moves (70% exploration) to ensure the network is exposed to sub-optimal, chaotic, and end-game positions.
• Normalization: The raw heuristic scores (which scale up to a theoretical mate score of 100,000) are clamped and normalized to [-1.0, 1.0] before being written to the JSON file.

Considerations for Using the Data

• The score field is strictly evaluated from the perspective of the player to move. If you are training a model to output an absolute score (e.g., White's advantage), you must invert the score when white_to_move is false.
• Escampe relies heavily on spatial constraints (the required band) and blocking logic. It is highly recommended to translate the bitboards into 6x6 spatial tensors to expose the orthogonal paths and band geometries to your neural network.

Citation

If you use this dataset, please consider linking back to this repository and crediting the Escampe AI project.