You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this model content.

Breakthrough MCVS - Zone Guided AI

Advanced Monte-Carlo Value Search (MCVS) engine for the game Breakthrough (8x8), powered by a displacement-based ABC Model and Weighted Adjacency Matrices with Hilbert-ordered Zone Guidance.

This implementation adapts the zone-guided MCVS framework to the simple but illustrative game Breakthrough, keeping the same neural architectures and zone-database design used by the chess reference implementation.

Core Idea

The engine uses:

  • Displacement-based ABC Model with homogeneous coordinates to represent piece displacements succinctly
  • Dynamic Weighted Adjacency Matrices W = A ⊙ S ⊙ F representing spatial, adjacency and feature similarity
  • Hilbert curve ordering for efficient neighborhood (zone) lookup and compression
  • A learned Zone Database that stores winning/losing/drawing position-pattern matrices and provides a k-NN based zone score
  • Zone Guidance integrated into PUCT (λ-PUCT) to bias MCTS toward favorable zones

The Breakthrough variant uses an internal 8×8 numpy board with lightweight move tuples (fr, fc, tr, tc). Policy outputs are flattened 4096-length move logits (from-square * 64 + to-square), and the value net predicts game outcome in [-1,1].

Files Overview

File Purpose
breakthrough_mcvs.py Full MCVS implementation for Breakthrough: game logic, ABC/WeightedMatrix classes, Policy/Value CNNs, Zone DB, MCVS & UCT searchers, self-play and training loop.
breakthrough_zone_db.npz Zone database file: stores Hilbert-ordered matrices for winning, losing, and draw zones used by zone guidance. Created/updated by breakthrough_mcvs.py.

Notes

  • The policy network maps a 1×64×64 weighted matrix tensor to a 4096-dimensional logits vector for flat move indexing.
  • The zone DB uses k-NN similarity (L1 normalized) across Hilbert-ordered matrices and returns a zone score Z ∈ [-1, 1].
  • breakthrough_mcvs.py includes a training loop that performs self-play data generation, incremental training, checkpointing (breakthrough_checkpoint.pt) and periodic MCVS vs UCT evaluation.

For implementation details, inspect breakthrough_mcvs.py. If you want a shorter quick-start, ask me to add a minimal README usage section with run commands and environment notes.

Downloads last month

-

Downloads are not tracked for this model. How to track
Video Preview
loading

Space using test1978/breakthrough-model 1