/RPC-Pytorch

Pytorch replication of the paper : Robust Predictable Control

Primary LanguagePythonMIT LicenseMIT

RPC-Pytorch

Pytorch replication of the paper : Robust Predictable Control

Run code

Run the original rpc agent with a KL constraint of 10 bits / timestep and evaluate with a noise factor of 3.

python train.py --agent 'RPC' --kl_constraint 10 --noise_factor 3

Run the rpc agent with a recurrent prior with a KL constraint of 1 bits / timestep and evaluate with a noise factor of 2.

python train.py --agent 'RRPC' --kl_constraint 10 --noise_factor 2

Command line arguments which can be set

  • device: use cuda or not
  • num_train_steps : total training time-steps
  • buffer_size maximum size of the data set
  • seq_len length of the sequence to sample for training the recurrent prior
  • lambda_init initial value for the dual parameter