Skip to main content

Hyperparameters

Guide to tuning hyperparameters for optimal training.

Common Parameters

ParameterPPO DefaultSAC DefaultDescription
learning_rate3e-43e-4Network learning rate
batch_size64256Training batch size
gamma0.990.99Discount factor
buffer_sizeN/A1MReplay buffer size

Tips

  1. Start with defaults - Our defaults work well for most dinosaur models
  2. Increase steps - More training usually helps
  3. Monitor rewards - Watch for plateaus
  4. Use GPU - Training is 10x faster
Coming Soon

Detailed hyperparameter tuning guide is under development.