The GNUbg Training Program

• First step - a benchmark for comparing two nets
  • The Benchmark
  • Disadvantages and objections
  • Establishing a benchmark - the rollout program
    • How to install
    • How to run
  • Creating the benchmark
• The Neural Net
  • Size of hidden layers
  • Hand Crafted Inputs
  • The Training Method
    • GNUbg-0.0 Training
    • Supervised Training
    • Use of Rollouts to Obtain Training Data
    • From Rollouts to 2ply
    • Poor Cube (to say nothing of 1ply)
• The Race Net
  • Analysis of race benchmark
  • Benchmark update using the large 12 point database
  • Error rates using the new benchmark
  • 2ply Vs. OSR racing
  • Conclusion
• The Crashed Net
  • Crashed benchmark positions
  • Analysis of crashed benchmark
• References

CHAPTERS

 Bad positions

• More on Ian Shaw positions

 Oystein explains how to use sagnu for windows

 Phase 1 contributions

• crashed positions
• Crashed doubles
• Race rollouts

 Status of rollouts in progress

• Rollout Team
• Contact Benchmark

DIRECTORIES