Training Specifications
Specify the following parameters in the ./configs/configs.txt file for specification of the training algorithm.
model
The neural network model to use in the muSim controller, can be [‘rnn’, ‘gru’]
hidden_size
The number of hidden units in the RNN/Feedforward layers of the muSim controller
mode
The mode of simulation can be [train, test, SFE, sensory_pert, neural_pert, musculo_properties]
Use ‘train’ for training and ‘test’ for testing the trained controller
(Remaining parameters are discussed in the perturbation modules section).
RL_algorithm = ‘SAC’
The RL algorithm can be either [SAC, DDPG, TD3] (Standard DRL algorithms)
cuda = True/False
Utilize GPU for training.
Other DRL specific parameters
Default values are recommended as they lead to succesfull training
gamma = 0.99
tau = 0.005
lr = 0.0003
alpha = 0.20
automatic_entropy_tuning = True
seed = 123456
policy_batch_size = 8
policy_replay_size = 4000
multi_policy_loss = True
batch_iters = 1
total_episodes = 1000000
condition_selection_strategy = “reward”
For the TD3 algorithm, the following parameters must be specified
target_noise = .2
target_noise_clip = .5
policy_delay = 2