OXERA: Grandmaster-Style Chess Policy Network

OXERA (Optimized Expert-level Engine with Residual Attention) is a high-fidelity chess policy network designed to bridge the gap between engine precision and human intuition. With 11.2 million parameters, OXERA is trained to replicate the decision-making processes of world-class players, specifically modeled after the gameplay of Magnus Carlsen and elite tournament participants (2500+ ELO).

🚀 Overview

Unlike traditional brute-force chess engines, OXERA operates as a Positional Intuition Engine. It does not merely calculate the highest mathematical advantage; instead, it predicts the most likely move a Grandmaster would make in a given position. This results in a highly aesthetic, human-like playing style that prioritizes dynamic piece activity and sophisticated positional understanding.

🧠 Model Architecture

Base Architecture: Residual Convolutional Neural Network (128 Filters, 6 Blocks).
Input Representation: 18-plane board encoding (Standard Maia/Lc0 format).
Parameters: 11,280,641.
Training Data:
- 700 MB of Lichess data.
- 250 MB of elite-level Lichess tournament data (Average ELO 2500+).

📈 Performance & Fidelity

OXERA excels in Move Prediction Accuracy, achieving professional-grade benchmarks in replicating elite human play:

Top-5 Accuracy: 96.3% (In 96 out of 100 positions, the Grandmaster's choice is within the model's top 5 candidates).
Top-1 Accuracy: ~46.5% (Matching the exact move of a world-class player in high-complexity positions).

The model demonstrates a profound understanding of:

Opening Nuances: High-fidelity replication of modern opening theory.
Strategic Transitions: Smooth handling of the transition from middle-game to endgame.
Prophylaxis: A strong tendency to anticipate and neutralize opponent plans before they manifest.

🛠️ Implementation & Usage

OXERA is a Policy-First network. While it provides exceptional move suggestions based on intuition, it is best utilized alongside a lightweight search algorithm (such as MCTS) to ensure tactical consistency in high-stakes environments.

Ideal Use Cases:

Interactive Analysis: Studying how a Grandmaster might approach a specific position.
Bot Development: Creating sophisticated chess personalities for platforms like Lichess.
Training Tool: Helping players understand positional concepts rather than just "engine lines."

📉 Training Methodology

The model was refined using advanced deep learning techniques to ensure stability and stylistic consistency:

Cosine Annealing: For optimal weight convergence.
EMA (Exponential Moving Average): To provide a balanced, stable version of the network's knowledge.
Expert Data Filtering: Only games from verified high-ELO sources were used to maintain the "Grandmaster" standard.

⚠️ License & Usage

This model is intended for research, educational, and analytical purposes.

jetbabareal
/

OXERA