#Parameters for uSim/nuSim

#Visualze the MuJoCo environment during training

visualize = False

#Print output statistics during training

verbose_training = True

###PATHS TO THE INPUT DATA/FILES###

###-------------------------------------------------------------------

## DO NOT change these paths if using the default paths for input data/files

#The path to the folder that contains the musculoskeletal model file

musculoskeletal_model_path = musculoskeletal_model/musculoskeletal_model.xml

#Path to the folder that contains intial pose (init_qpos.npy and init_qvel.npy) files

initial_pose_path = initial_pose

#Path to the folder that contains the experimental kinematics data

kinematics_path = kinematics_data

#Path to the folder that contains the experimental neural data

nusim_data_path = nusim_neural_data

#Path to the folder that contains the experimental stimulus data

stimulus_data_path = stimulus_data

### PATHS FOR SAVING THE OUTPUT / TEST DATA###

### DO NOT change these paths if using the default paths for the saved output/test data

###--------------------------------------------------

#Save the agent networks after save_iter 

save_iter = 100

#Path to the root directory

root_dir = "."

#Path to save the agent's neural networks

checkpoint_folder = "./checkpoint"

#Path for saving the statististics for training

statistics_folder = "training_statistics"

#Save name for the agent's networks

checkpoint_file = "agent_networks"

#Save name for saving the test data

test_data_filename = "test_data"

#Load the saved networks from the previous session for further training

load_saved_nets_for_training = False

### Kinematics Preprocessing Parameters

###----------------------------------------------------------------------

#Kinematics preprocessing for simulation

#Adjustment instructions:

#The timestep for the simulation: Keep 0 for default simulation timestep

sim_dt = 0   # in seconds

#The frames/timepoints for which the same action should be repeated during training of the agent

#For finer movements user smaller frame_repeat, but it will also increase the training time

frame_repeat = 5

#Number of fixedsteps in the beginning of the simulation. The target will remain at kinematic[timestep=0] for n_fixedsteps

#If a good initial position is found using CMA-ES / IK Optimization, n_fixedsteps = 25 is a good estimate. Otherwise increase

#if the starting reward does not increase with the training iterations.

n_fixedsteps = 25

#Timestep limit is max number of timesteps after which the episode will terminate.

#Multiple cycles of the same condition will be simulated if the timestep_limit > number of timsteps for that condition.

timestep_limit = 8000

#Adjusts/scales the length of the trajectory

#Should be the same as num_markers/targets

trajectory_scaling = [26.3157894737]

#Adjusts the starting point of the kinematics trajectory

#Should be the same as num_markers/targets, num_coords=3

center = [[0.06, 0.083, 0]]

###-----------------------------------------------------------------------------------

###Sensory Feedback Processing Parameters --------------------------------------------

#Specifies the sensory feedback to the agent/network

#True, if this feedback should be included in state feedback to the agent's network/controller

#False, if this feedback should not be included in the state feedback to the agent's network/controller

#Stimulus feedback consists of provided experimental stimulus data

stimulus_feedback = False

#Proprioceptive feedback consists of muscle lengths and velocities

proprioceptive_feedback = True

#Muscle forces consist of appled muscle forces 

muscle_forces = False

#Joint feedback consists of joint positions and velocities

joint_feedback = False 

#Visual feedback consists of x/y/z coordinates of the specified bodies in the model

#If visual_feedback is True, specify the names of the bodies from musculoskeletal_model.xml for which the feedback should be included

visual_feedback = False 

#Append the musculo bodies from which visual feedback should be included

#This list can also consist of targets/markers

#Append targetn-1 for visual feedback from targets/markers in the kinematics.pkl file

#'target0' corresponds to the visual feedback from the first target/marker, target1 to the second target/marker and so on

visual_feedback_bodies = [hand, target0] 

#Specify the names of the bodies as tuples(separated by ; with no spaces) for which the visual distance should be included in the feedback

#Leave blank if the visual distance is not to be included in the feedback

#Visual distance between the bodies will be included

#e.g visual_distance_bodies = [[hand;target0], [elbow;target0]] will include the distance between the hand/elbow and first marker in sensory feedback

visual_distance_bodies = [[hand;target0]] 

#Specify the names of the bodies for which the visual velocity should be included in the feedback

#Leave blank if the visual velocity is not to be included in the feedback

#Appends the absolute musculo body velocity, e.g. visual_velocity = [hand, target0] 

#will include the xyz velocities of hand and target0

visual_velocity = []

#Specify the delay in the sensory feedback in terms of the timepoints

sensory_delay_timepoints = 0

### -----------------------------------------------------------------------------------

###Specifications for Regularizations with the policy network

#Specify the weighting with various neural regularizations used in uSim/nuSim

#weighting with loss for enforcing simple neural dynamics for uSim/nuSim

alpha_usim = 0.1

#weighting with loss for minimizing the neural activations for uSim/nuSim

beta_usim = 0.01 

#weighting with loss for minimizing the synaptic weights for uSim/nuSim

gamma_usim = 0.001

#weighting with loss for nuSim constraining a sub-population of RNN units to experimentally recorded neurons for nuSim

zeta_nusim = 0

### --------------------------------------------------------------------------------------

### SAC TRAINING ###

#The neural network model to use in the agent, can be ['rnn', 'gru']

model = rnn

#The number of hidden units in the layers of the agent's neural network

hidden_size = 256

#The mode of simulation can be [train, test, SFE, sensory_pert, neural_pert, musculo_properties]

mode = "train"

#DRL specific parameters.

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"

cuda = True