All articles

Do people leave a trace when they play chess?

2026

My bachelor's thesis, on characterising chess stylometry with machine learning and composite images, does not propose another engine or look for the best move. It asks something narrower, whether a person leaves a recognisable trace when playing.

Approach

The work uses Lichess games and has one main restriction, which is not using Stockfish or engine evaluations as an input signal. The system only sees what can be observed directly in the game, as board images and decision-time maps.

These train a Siamese network with metric learning and triplet loss, which produces game embeddings to compare similarity between players.

System

The pipeline covers game extraction and filtering, image generation, board-perspective normalisation, training, and inference based on a memory bank. The bank keeps several games per player instead of reducing everything to an average, which makes it possible to retrieve close examples and study the model's confusions.

Results

The results are modest, but they show signal.

EvaluationResultReference
Identification, 10 playersTop-5 of 79.75% to 82.25%50% for random Top-5
Identification, 200 playersTop-1 of 2.05%0.5% for random Top-1
Raw EloMAE of 329.0Predictions compressed to the centre
Calibrated EloMAE of 260.6Out-of-fold linear calibration
Elo over 20 gamesMAE of about 224Best calibrated method

With 200 players exact identification drops sharply, as expected, though it stays about four times above chance. The model was not trained to estimate Elo, but the embedding does carry information about playing level.

Conclusion

The system is not a robust authorship identifier and does not try to compete with engine-based methods. It is a complete first implementation of a visual and temporal approach to comparing how people play.

The most valuable outcome is confirming that there is usable signal and understanding the problem better, because style, level, colour, game phase, clock and repertoire are all mixed together. Separating those factors remains future work.