Deep Neural Networks for Double Dummy at Bridge
summary
A deep neural network predicts double dummy trick counts using TensorFlow with convolutional layers and ResNet-style skip connections and trains on millions of sampled deals. The model and experiments replace slow exact solvers to run 750–1000 deals/sec versus ~7/sec for exact search, yielding about ~70% exact accuracy (better than prior NNs and often above human experts except in no-trump) and reducing errors mostly to one trick. Code and pretrained models are available on GitHub and S3.
AI-written summary · about
chapters (15)
- 0:00Bridge basics and motivation
- 3:19Talk outline and bridge basics
- 5:11Current bridge AI challenges
- 7:22Neural network input for bridge
- 9:11Baseline models and metrics
- 12:41Improving neural network performance
- 14:26Convolutional neural network architecture
- 18:21Adding ResNet skip connections
- 19:28Results versus human experts
- 20:32Speed versus exact solvers
- 22:44Open questions and future work
- 23:29Pre-trained model for predictions
- 24:52Project lessons learned
- 27:42Extending model for later stages
- 28:45Q&A
related talks
59mDeep Neural Networks for Video Applications
25mBridging Gaps: Classic Approaches Beyond Deep Learning Constraints
01h 30mDeep Learning Techniques for Monocular Depth Estimation
38mThe Joy of Designing Deep Neural Networks
37mDeep Recommender Systems
48mDynamic Neural Networks with DyNet
41mDeep Learning to Solve Challenging Problems
56mDeep Fakes 2.0 - How Neural Networks are Changing our World