Pytorch replication of the paper : Robust Predictable Control
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
device
: use cuda or notnum_train_steps
: total training time-stepsbuffer_size
maximum size of the data setseq_len
length of the sequence to sample for training the recurrent priorlambda_init
initial value for the dual parameter