test1978/breakthrough-model

Breakthrough MCVS - Zone Guided AI

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

This repository implements a complete zone-guided reinforcement learning system, including self-play training, neural networks, and comparative tournaments against classic UCT.

Core Idea

The engine uses:

• Displacement-based ABC Model with homogeneous coordinates
• Dynamic Weighted Adjacency Matrices W = A ⊙ S ⊙ F
• Hilbert curve ordering for efficient zone retrieval
• A learned Zone Database that stores winning/losing position patterns
• Zone Guidance (λ-PUCT) to bias search toward promising zones

For more information please refer to the paper at: https://doi.org/10.13140/RG.2.2.18795.09764

Files Overview

File	Purpose
breakthrough_mcvs.py	Main implementation: game logic, ABC model, Zone Database, MCVS, neural networks, incremental training
mcvs_vs_uct.py	200-game tournament between MCVS and UCT with detailed logging and online learning
abc_model.py	Displacement-based ABC Model
matrix_model.py	Computes the weighted adjacency matrix
breakthrough_zone_db.npz	Learned Zone Database
breakthrough_checkpoint.pt	Saved Policy & Value neural network weights
inspect_npz.py	Utility to inspect the zone database

Requirements

How to use:

A. Incremental Training

python breakthrough_mcvs.py

This runs continuous self-play + training:

1. Generates games using MCVS
2. Trains the neural networks
3. Updates and saves the Zone Database
4. Fully incremental (you can stop and resume anytime)

B. Tournament with Online Learning

This script runs a 200-game tournament while the AI learns.

With Neural Networks (Full version – Online Learning):

bash: python mcvs_vs_uct.py

What happens:

1. MCVS plays against classic UCT (alternating sides)
2. Neural Policy and Value networks learn online after every game
3. Zone Database is updated and saved after each game
4. Zone guidance turns on automatically after game 1
5. Creates detailed logs:
6. breakthrough_full_results.txt — tournament summary
7. move_log.txt — per-move statistics
8. learning_log.txt — training progress

Without Neural Networks (Zone-only Ablation)

To run the faster ablation version:

1. Open mcvs_vs_uct.py
2. Change this line near the bottom:

ablation_no_nets=False # ← Change to True

pip install torch numpy