behaviors:
  Kyle:
    trainer_type: sac
    hyperparameters:
      learning_rate: 0.0003
      learning_rate_schedule: linear
      batch_size: 1000
      buffer_size: 1000000
      tau: 0.005
      steps_per_update: 30.0
      init_entcoef: 1.0
      reward_signal_steps_per_update: 30.0
    network_settings:
      normalize: true
      hidden_units: 256
      num_layers: 3
      vis_encode_type: simple
      memory:
        sequence_length: 1000
        memory_size: 256
    reward_signals:
      extrinsic:
        gamma: 0.995
        strength: 1.0
      curiosity:
        strength: 0.1
        gamma: 0.995
        network_settings:
          hidden_units: 256
          learning_rate: 0.0001
    keep_checkpoints: 40
    checkpoint_interval: 500000
    max_steps: 20000000
    time_horizon: 1000
    summary_freq: 10000