Otter: A New Human Chess AI with Time-Aware Modeling
A new chess AI named Otter, featuring 15.3 million parameters, has been developed by researchers to forecast human moves by viewing gameplay as a sequential, time-sensitive process instead of analyzing each position independently. This model integrates two key conditioning elements: a move history encoder that utilizes the last 20 moves to reflect opening choices, positional shifts, and player tendencies, along with a time control module that adjusts predictions according to clock constraints. Trained on 6.1 billion positions from 117 million rapid games on Lichess over 30 days using a single T4 GPU, Otter achieves a top-1 accuracy of 55.23% and a top-5 accuracy of 90.95%, outperforming the previous leading human chess model, Maia 2. Its accuracy reaches 57.38% in the 1900-1999 Elo range. The findings suggest that a time-aware approach enhances prediction accuracy. The research paper can be found on arXiv with the identifier 2608.05206.
Key facts
- Otter is a 15.3M-parameter human chess AI
- It predicts human move selection by modeling play as a time-aware, sequential process
- It uses a move history encoder conditioning on the last 20 moves
- It includes a time control module that modulates predictions based on clock pressure
- Trained on 6.1 billion positions from 117 million Lichess rapid games
- Training took 30 days on a single T4 GPU
- Achieves 55.23% top-1 and 90.95% top-5 move-prediction accuracy
- Surpasses Maia 2 with fewer parameters and less training data
- Accuracy peaks at 57.38% in the 1900-1999 Elo bracket
Entities
Institutions
- arXiv
- Lichess