|
|
--- |
|
|
license: mit |
|
|
pipeline_tag: reinforcement-learning |
|
|
tags: |
|
|
- chess |
|
|
- engine |
|
|
datasets: |
|
|
- jetbabareal/veri_txt |
|
|
--- |
|
|
|
|
|
# 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. |
|
|
|
|
|
--- |
|
|
|
|
|
### Tags: |
|
|
`Chess AI` `Magnus Carlsen` `Grandmaster Intuition` `Policy Network` `PyTorch` `Leela Chess Zero` `Human-like AI` |
