scone_walk_h2190.yaml

tonic:
  after_training: ''
  header: "import deprl, gym, sconegym"
  agent: "deprl.custom_agents.dep_factory(3, deprl.custom_mpo_torch.TunedMPO())(replay=deprl.custom_replay_buffers.AdaptiveEnergyBuffer(return_steps=1,
    batch_size=256, steps_between_batches=1000, batch_iterations=30, steps_before_batches=2e5,
    num_acts=90))"
  before_training: ''
  checkpoint: "last"
  environment: "deprl.environments.Gym('sconewalk_h2190-v1', scaled_actions=False)"
  full_save: 1
  name: "sconewalk_h2190_v1"
  resume: true
  seed: 0
  parallel: 20
  sequential: 10
  test_environment: null
  trainer: "deprl.custom_trainer.Trainer(steps=int(5e8), epoch_steps=int(2e5), save_steps=int(1e6))"

working_dir: "IGNORED_FOR_HYFYDY"

env_args:
  clip_actions: true
  grf_coeff: -0.17281
  joint_limit_coeff: -0.1307
  nmuscle_coeff: -1.57929
  smooth_coeff: -0.097
  vel_coeff: 10
  step_size: 0.025
  init_activations_mean: 0.01
  init_activations_std: 0 # if 0: deterministic

mpo_args:
  hidden_size: 1024
  lr_actor: 3.53e-05
  lr_critic: 6.081e-05
  lr_dual: 0.00213

DEP:
  bias_rate: 0.002
  buffer_size: 200
  intervention_length: 8
  intervention_proba: 0.00371
  kappa: 1000
  normalization: "independent"
  q_norm_selector: "l2"
  regularization: 32
  s4avg: 2
  sensor_delay: 1
  tau: 40
  test_episode_every: 3
  time_dist: 5
  with_learning: true