Mini-Chess Zero

Asymmetric TD Learning: A biologically-inspired approach to stable deep reinforcement learning

🎯 Results

Metric	Before	After
Win Rate	3%	87.5%
Q-Value	70,615,864	9.29
Status	Diverging	Converged

Key Achievement: 1,380× reduction in Q-value explosion through novel stabilization techniques.

---

🧠 Novel Contribution: Asymmetric TD Learning

Inspired by dopamine neuron asymmetry in biological brains:

# Positive TD errors: Learn cautiously (weight = 0.5)
# Negative TD errors: Learn aggressively (weight = 1.5)
weights = torch.where(td_errors > 0, 0.5, 1.5)
loss = (weights * huber_loss(predicted, target)).mean()

This prevents overoptimistic value estimates while enabling rapid learning from mistakes.

---

📦 Installation

# Clone repository
git clone https://github.com/your-username/mini-chess-zero.git
cd mini-chess-zero

# Create virtual environment
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Install the asymmetric-td package locally
pip install -e .

---

🚀 Quick Start

Train the Agent

python train_monitor.py

Play Against AI

python play_human.py

Run Web App

cd webapp && python api.py
# Open http://localhost:8000

Run Trading Backtest

cd trading && python backtest.py

---

📁 Project Structure

mini_chess_rl/
├── src/                    # Core RL (87.5% win rate)
│   ├── agent.py            # Stabilized DQN
│   ├── environment.py      # 5x5 Mini-Chess
│   └── replay_buffer.py
├── asymmetric_td/          # Python Package
│   ├── losses/             # AsymmetricTDLoss
│   ├── agents/             # StableDQN
│   └── utils/              # Gradient, soft update
├── webapp/                 # Chess Web App
├── trading/                # Trading Adaptation (+24% return)
├── research/               # Paper, ablation study
└── models/                 # Trained models

---

📊 Ablation Study

Configuration	Q-Value	Status
FULL	9.29	✅ Stable
NO_ATD	9.23	✅ Stable
NO_Q_CLIP	10.82	⚠️ Drifting
VANILLA	12,822	💥 Exploded

The combined stabilization stack is essential—removing all techniques causes catastrophic failure.

---

🎮 Features

• Asymmetric TD Learning - Novel biologically-inspired loss function
• Gradient Clipping - Prevents explosion (max_norm=10)
• Polyak Soft Updates - Smooth target updates (τ=0.005)
• Q-Value Clipping - Bounds estimates ([-10, 10])
• Reward Centering - Reduces variance
• Action Masking - Legal move enforcement

---

📈 Neural Network Analysis

The trained agent learned classical chess piece values:

Piece	Learned Q-Value
King	+3.18
Queen	+2.97
Rook	+2.64
Bishop	+2.26
Knight	+1.98
Pawn	+1.86

---

📄 Research Paper

See research/paper/paper_draft.md for the complete paper:

• Title: "Asymmetric Temporal Difference Learning: Biologically-Inspired Stabilization for Deep RL"
• Key Result: 1,380× reduction in Q-value magnitude

---

🔧 Configuration

Key hyperparameters in src/agent.py:

Parameter	Value	Description
GAMMA	0.95	Discount factor
LEARNING_RATE	0.0001	Adam LR
TAU	0.005	Soft update rate
GRAD_CLIP	10.0	Gradient clipping
ATD_WEIGHTS	(0.5, 1.5)	Asymmetric weights

---

📜 License

MIT License - see LICENSE for details.

---

🙏 Acknowledgments

• Gardner Mini-Chess variant
• PyTorch team
• Schultz (1997) for dopamine asymmetry research