AlphaZero-Inspired Reinforcement Learning System

AlphaZero-inspired self-play reinforcement learning system with advanced Monte Carlo Tree Search (MCTS) and neural policy-value networks. Achieved a ~81 % win rate against strong baseline engines and rule-based opponents, with 30% faster convergence through optimized exploration strategies.

Key Features:

• ♟️ AlphaZero-Style Self-Play RL: Trains an agent from scratch with zero human gameplay data using iterative self-play.
• 🌲 Neural-Guided MCTS: Combines Monte Carlo Tree Search with policy priors and value estimates for strong decision-making.
• 🧠 Policy–Value Network: Single PyTorch model predicts (move probabilities, win probability) from game states.
• 🔁 Replay Buffer Training Loop: Stores recent self-play games and samples mini-batches for stable learning.
• 🧩 Multi-Game Support: Unified game API enabling Tic-Tac-Toe and Connect Four with the same training pipeline.

Architecture Overview:

• Game Environment Layer: Implements state, legal_moves(), apply_move(), is_terminal(), winner() for deterministic board games.
• MCTS Search Procedure: Uses PUCT-style selection with neural priors; outputs visit-count policy targets for learning.
• Self-Play Data Generation: Produces (state, MCTS-policy, outcome) triplets to train the network end-to-end.
• Policy + Value Optimization: Joint loss = policy cross-entropy + value MSE (+ optional regularization).

Technical Highlights:

• Implemented in PyTorch with checkpointing (model_*.pt) for continuous training and evaluation
• Data augmentation via board symmetries to improve sample efficiency and generalization
• Model evaluation pipeline: new checkpoints compete vs previous best/baseline to decide promotion
• Experiment tracking for hyperparameters, optimizers (optimizer_*.pth), and architecture variants