renegade-net-35

This is the neural network used for evaluating positions in the Renegade chess engine. The architecture is (768x14hm -> 1600)x2 -> 1x8, which allows for extremely fast CPU inference. The neural network is uploaded both in safetensors and in raw binary format, as well as a reference script for inference. It is only intended for evaluating 'quiet' positions (where there are no possible winning captures), otherwise capture chains should be resolved before inference.

Net architecture

• Perspective net
• 14 input buckets based on the position of the king
• 8 output buckets based on the remaining piece count
• Horizontal mirroring of pieces

Usage

network = ChessNeuralNet("renegade-net-35.safetensors")
score = network.evaluate_position("rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1")
print(f"Evaluation for the starting position: {round(score)} units")

Evaluation for the starting position: 107 units

Scores are relative to the side to move: positive values correspond to an advantage, while negative values imply a losing position. Outputs in the high triple digits and above typically represent a won game, whereas near-zero values indicate a drawish position.

Other information

• Parameter count: 17.2 million
• Training data size: 6.30 billion positions
• Training time: 62 hours (mediocre consumer GPU)