control.yaml

# Environment config
env:
  type: 'classic_control'
  # Classic control env name
  name : 'CartPole-v0'
  # seed
  seed: 543
  # solution rewards
  env_solution: 195

# Agent config
agent:
  # type of agent
  agent_type: 'vpg'
  # model type
  model_type: 'mlp'
  # Learning rate for the policy network
  policy_lr : 0.01
  # Learning rate for the agent network
  value_lr : 0.001
  # optimizer name
  opt_name: 'adam'
  # norm of gradient clipping, leave empty for no clipping
  grad_clip:
  # Bellman equation reward discount
  gamma : 0.99
  # Estimated GAE with TD(lambda)
  lambda: 0.97
  # vpg scaling:
  vpg_scaling: 1.0
  # value loss scaling
  value_scaling: 1.0
  # entropy loss scaling
  entropy_scaling: 0.005
  # normalize rewards:
  reward_norm: true
  # input shape
  input_shape : [4]
  # state size := [state_size] + [input_shape]
  state_len: 1
  # action_size
  action_size: 2
  # transform the input
  input_transforms:

train:
  # Number of training episodes
  n_train_episodes : 100
  # Max steps in each episode
  max_steps : 1000
  # model location
  model_dest: /data/experiments/agent-of-control/02-12-2020-CartPole-v0-vpg-gae
  # save model every save_model steps
  save_model: 20

test:
  # Number of testing episodes
  n_test_episodes : 5
  # Max steps in each episode
  max_steps : 1000
  # path where to save played video
  state_dest: /data/experiments/agent-of-control/02-12-2020-CartPole-v0-vpg-gae/states