default_config_pendulum_ppo_opt.yaml

env_name: !!str Pendulum-v0
device: !!str cuda                          # torch device (cuda:0 or cpu)
render_env: !!bool False                    # render environment

agents:
  ppo:
    train_episodes: !!int 10000             # maximum number of episodes to optimize
    test_episodes: !!int 10                 # maximum number of episodes to evaluate
    init_episodes: !!int 0                  # number of episodes to fill the replay buffer
    update_episodes: !!float 13             # update policy every x episodes (can be float)
    ppo_epochs: !!int 170                   # update policy for x epochs
    gamma: !!float 0.952                    # discount factor
    lr: !!float 0.000919                    # learning rate
    vf_coef: !!float 0.204                  # value function coefficient (see PPO paper)
    ent_coef: !!float 0.00434               # entropy coefficient (see PPO paper)
    eps_clip: !!float 0.539                 # trust region size (see PPO paper)
    rb_size: !!int 1000000                  # size of the replay buffer
    same_action_num: !!int 1                # how often to perform the same action subsequently
    activation_fn: !!str tanh               # activation function for actor/critic ('tanh', 'relu', 'leakyrelu' or 'prelu')
    hidden_size: !!int 92                   # size of the actor/critic hidden layer
    hidden_layer: !!int 2                   # number of hidden layers
    action_std: !!float 1.57                # action noise standard deviation
    print_rate: 1                           # update rate of avg meters
    early_out_num: !!int 50                 # based on how many training episodes shall an early out happen
    early_out_virtual_diff: !!float 0.0291  # performance difference for an early out for virtual envs

envs:
  Pendulum-v0:
    solved_reward: !!float -300             # used for early out in RL agent training
    max_steps: !!int 200                    # maximum number of steps per episode
    activation_fn: !!str leakyrelu          # activation function of the virtual environment
    hidden_size: 32                         # size of the hidden layer of the virtual environment
    hidden_layer: !!int 2                   # number of hidden layers of the virtual environment
    info_dim: !!int 0                       # additional information dimension from step function
    reward_env_type: !!int 2                # type of reward shaping function