/aegis-rnd

Random network distillation in Aegis

Primary LanguagePython

Aegis Random Network Distillation node

Uses https://github.com/tehZevo/rnd for the RND implementation

Environment variables

  • PORT: the port to listen on
  • INPUT_SIZE: input (observation) vector size
  • OUTPUT_SIZE: RND model output vector size
  • PREDICTOR_HIDDEN: hidden layer sizes for the predictor network, separated by spaces e.g., 128 64 (defaults to 32)
  • TARGET_HIDDEN: as above, but for the target network, defaults to 32 32
  • ACTIVATION: hidden layer activation function for both models, defaults to "swish"
  • LR: learning rate, defaults to 1e-3
  • BUFFER_SIZE: training buffer size, defaults to 10000
  • BATCH_SIZE: training batch size, defaults to 32
  • SAVE_PATH: path to save/load models from, defaults to models
  • AUTOSAVE_STEPS: save every steps, defaults to 1000