Research map / Architectures and efficiency / Attention and transformers
Fast and Knowledge-Free Deep Learning for General Game Playing (Student Abstract)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Adapts AlphaZero to general game playing in the Regular Boardgames system using value-only attention networks trained on MCTS-vs-MCTS data, removing action-space and board-topology assumptions.
Problem
Neural network game agents in General Game Playing typically need the action space and board topology extracted from game rules and require long self-play training; the paper targets fast, knowledge-free model building.
Approach
Replace AlphaZero's policy-plus-convolutional network with a value-only attention network over tile embeddings, use plain UCT (no PUCT priors), and generate training data from UCT MCTS playouts instead of self-play, so no action-space or board-topology knowledge is needed.
Key results
- Attention NN achieves 75% win rate vs 1x UCT and 47% vs 10x UCT on average on ordered boards (600 iterations), versus 69% and 50% for the convolutional NN.
- With permuted boards, attention NN averages 64% vs 1x UCT and 39% vs 10x UCT, while the CNN averages 56% and 38%.
- After very short training (under 3% of the 2h dataset, about 1,000 MCTS plays), Breakthrough 6x6 with attention reaches 74% vs UCT, and games like Reversi and English Draughts reach 28% and 54%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use attention layers instead of convolutional layers | AlphaZero-style value network · Total average, ordered board vs UCT 1× | win rate: 69% -> 75% | Helped |
| Use attention layers instead of convolutional layers | AlphaZero-style value network · Total average, ordered board vs UCT 10× | win rate: 50% -> 47% | Hurt |
| Permute the board tile order (ablation) | Attention NN · Total average vs UCT 1× | win rate: 75% -> 64% | Hurt |
| Permute the board tile order (ablation) | Convolutional NN · Total average vs UCT 1× | win rate: 69% -> 56% | Hurt |
| Increase the training dataset from 200 to 1,000 MCTS plays (ablation) | Attention NN · Breakthrough (6×6) vs UCT 1× | win rate: 4% -> 74% | Helped |
| Increase the training dataset from 200 to 1,000 MCTS plays (ablation) | Attention NN · English Draughts vs UCT 1× | win rate: 31% -> 54% | Helped |
| Increase the training dataset from 200 to 1,000 MCTS plays (ablation) | Attention NN · Reversi vs UCT 1× | win rate: 25% -> 28% | Mixed |
| Use attention layers instead of convolutional layers | AlphaZero-style value network · Across games and board sizes | win rate | Mixed |
Surprising findings
- Optimized MCTS playouts in the RBG reasoner are much faster than NN evaluation, so time-based comparisons favor pure MCTS.
- Convolutional NN outperforms attention NN on several smaller games, so attention is not uniformly better; trend favors attention only for larger games.
Limitations
- Experiments focus on budget training; longer training settings may change the landscape.
- Attention NN does not uniformly beat CNN; results vary strongly by game and size.
- Limited to two-player zero-sum perfect-information games.
Experimental setup
Datasets: MCTS self-play-free plays generated by UCT agents on Regular Boardgames games (Breakthrough, Connect Four, English Draug
Benchmarks: Regular Boardgames games: Breakthrough, Connect Four, English Draughts, Fox and Hounds, Gomoku, Hex, Pentago, Reversi, T
Metrics: win rate vs UCT with 1x and 10x iterations, 95% confidence i
Baselines: UCT MCTS Convolutional NN value network
Builds on: AlphaZero UCT Regular Boardgames PUCT
Closest papers by meaning
- Enhancing Chess Reinforcement Learning with Graph Representation (NeurIPS 2024 · similarity 0.89)
- PokerBench: Training Large Language Models to Become Professional Poker Players (AAAI 2025 · similarity 0.85)
- Can Large Language Models Master Complex Card Games? (NeurIPS 2025 · similarity 0.84)
- Mastering Board Games by External and Internal Planning with Language Models (ICML 2025 · similarity 0.84)
- NfgTransformer: Equivariant Representation Learning for Normal-form Games (ICLR 2024 · similarity 0.84)
- Diffusion Models Are Real-Time Game Engines (ICLR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.