Otter: A Time-Aware, History-Conditioned Human Chess AI
Otter is a 15.3M-parameter chess AI using history and time signals to predict human moves with 55.23% accuracy.
Key Findings
Methodology
Otter predicts chess moves using a history encoder and a time control module. The history encoder processes the last 20 moves using a Transformer, while the time control module adjusts predictions based on clock pressure.
Key Results
- Otter achieves 57.38% top-1 accuracy in the 1900-1999 Elo bracket, outperforming Maia 2.
- Across all Elo brackets, Otter's top-1 prediction accuracy is 55.23%, with top-5 accuracy at 90.95%.
- By combining history and time signals, Otter achieves higher human move prediction accuracy with a smaller model.
Significance
Otter's research impacts academia and industry by addressing the shortcomings of traditional models that ignore time pressure and historical behavior, offering new possibilities for personalized coaching and cheat detection.
Technical Contribution
Otter integrates time and history signals into a single model, providing new theoretical guarantees and engineering possibilities compared to existing state-of-the-art methods.
Novelty
Otter is the first to combine time and history signals for move prediction, offering more accurate predictions compared to Maia 2.
Limitations
- Otter's accuracy decreases under extreme time pressure as the model fails to fully capture the complexity of rapid decision-making.
- The model performs poorly in lower Elo brackets due to the more random nature of move selection by these players.
Future Work
Future work could include improving the time control module to better handle extreme time pressure and expanding the model to cover a wider range of board games.
AI Executive Summary
Otter is a novel chess AI designed to predict human moves by using time-aware and history-conditioned signals. Traditional chess move prediction models often overlook the sequential nature of games and the impact of time pressure on decision-making. Otter successfully improves prediction accuracy by integrating a move history encoder and a time control module. Experimental results show that Otter performs exceptionally well across various Elo brackets, particularly achieving 57.38% top-1 accuracy in the 1900-1999 Elo bracket. This research holds significant academic and practical implications, offering new possibilities for personalized coaching and cheat detection. However, Otter's performance under extreme time pressure remains a challenge. Future research directions include optimizing the time control module and expanding the model's applicability.
Deep Analysis
Background
With the advancement of AI technology, chess engines have surpassed human players. However, predicting human moves remains challenging due to the influence of historical behavior and time pressure on decision-making. Maia and Maia 2 were early attempts to predict human moves but ignored the sequential nature and time factors of games.
Core Problem
Traditional move prediction models often treat each position as an independent event, ignoring the influence of historical behavior and time pressure on decision-making. This approach fails to accurately predict human moves as it does not capture player momentum and game trends.
Innovation
Otter addresses this issue by combining a move history encoder and a time control module. The history encoder uses a Transformer to process the last 20 moves, capturing player opening preferences and game trends. The time control module adjusts predictions based on clock pressure.
Methodology
- �� History encoder processes the last 20 moves using a Transformer.
- �� Time control module adjusts predictions based on clock pressure.
- �� Combines both signals to improve prediction accuracy.
Experiments
Otter is trained on 117 million Lichess rapid games using 6.1 billion positions. Experimental design includes performance tests across different Elo brackets and comparisons with Maia 2.
Results
Experimental results show that Otter achieves 57.38% top-1 accuracy in the 1900-1999 Elo bracket, outperforming Maia 2. Overall top-1 prediction accuracy is 55.23%, with top-5 accuracy at 90.95%.
Applications
Otter can be used for personalized coaching and cheat detection. It provides more accurate human move predictions, aiding in the development of more human-like AI opponents.
Limitations & Outlook
Otter's performance under extreme time pressure remains a challenge. The model performs poorly in lower Elo brackets due to the more random nature of move selection by these players.
Plain Language Accessible to non-experts
Imagine a kitchen where a chef needs to complete a dish within a limited time. Traditional AI is like a chef who only focuses on the current ingredients, while Otter is like a chef who considers the entire cooking process and time pressure. It not only focuses on current ingredients but also considers previous steps and remaining time, making decisions that align more with human habits.
ELI14 Explained like you're 14
Hey, imagine you're playing a super complex board game. Usually, AI only focuses on the pieces on the board, but Otter is smarter! It considers every move you've made before and also pays attention to the time you have left. It's like a super smart opponent that knows what you're thinking and can predict your next move. Cool, right?
Glossary
Transformer
A neural network architecture used for processing sequential data, adept at capturing long-range dependencies.
Used in the move history encoder to process the last 20 moves.
Elo Rating
A rating system used to evaluate the skill level of chess players; higher numbers indicate higher skill.
Used to categorize player skill levels to adjust the prediction model.
Cross-Attention
An attention mechanism allowing the model to exchange information between different data streams.
Used to combine board and history information to improve prediction accuracy.
Time Control
Time constraints in a chess game that affect the player's decision-making process.
Used to adjust predictions to account for clock pressure.
Dropout
A technique to prevent overfitting by randomly dropping connections in a neural network.
Used in CNN layers to improve the model's generalization ability.
Open Questions Unanswered questions from this research
- 1 How to improve prediction accuracy under extreme time pressure? Current models fail to fully capture the complexity of rapid decision-making.
- 2 How to expand the model to cover more board games? Current models are limited to chess.
Applications
Immediate Applications
Personalized Coaching
Provides targeted coaching advice by predicting moves more accurately, improving learning efficiency.
Cheat Detection
Helps detect potential cheating by identifying abnormal move choices.
Long-term Vision
Cross-Board Game AI Development
Expands the model to cover more board games, promoting comprehensive development of board game AI.
Abstract
Otter is a 15.3M-parameter human chess AI that predicts human move selection by modeling play as a time-aware, sequential process rather than treating each position in isolation. It combines two conditioning signals: (1) a move history encoder that conditions predictions on the last 20 moves, capturing opening preferences, positional drift, and intra-game behavioral tendencies; and (2) a time control module that modulates predictions based on clock pressure. Otter is trained on 6.1 billion positions from 117 million Lichess rapid games over 30 days on a single T4 GPU. Otter achieves 55.23% top-1 and 90.95% top-5 move-prediction accuracy, surpassing the prior state-of-the-art human chess model, Maia 2, with far fewer parameters and less training data. Across 11 Elo brackets (<1100 to >=2000), accuracy peaks at 57.38% in the 1900-1999 bracket. These results show that modeling chess as a time-aware, sequential activity yields more human-accurate move prediction than position-only approaches, using a smaller model. Code, trained models, and complete training logs are publicly released.