Views
No views yet
1# This file is used to configure logging and agents behaviour.
2#
3# The first part consists of Wandb info used to log experiments.
4# Changing it adjusts the way logging is stored and displayed.
5#
6# The second part (config) is used to change hyperparameter settings of agents.
7# Changing it adjusts the way agents behave and learn.
8project: "DM Control Reach"
9name: "PPO"
10dir: "logs"
11notes: "Training DM Control Reach using PPO"
12monitor_gym: "False"
13config:
14 # Environment, logging and saving control
15 environment: "dm_control/reacher-hard-v0" # Environment to use
16 algorithm: "PPO Continuous" # What kind of algorithm to use?
17 save_dir: "models" # Where to save model?
18 save_name: "ppo_control_reach" # Model name
19 save_interval: 50 # How many previous episodes will be used to calculate mean reward?
20 total_steps: 1_000_000 # For how many steps will the agent train?
21 # Algorithm hyperparameters
22 gamma: 0.99 # Discount factor for future rewards
23 lambda: 0.95 # GAE tradeoff parameter
24 ppo_epochs: 10 # How many epochs to train on each batch?
25 rollout_length: 2048 # How many steps to collect before updating?
26 batch_size: 64 # How many steps are in each batch?
27 clip_epsilon: 0.2 # How much is the policy clipped?
28 learning_rate_actor: 0.0003 # Learning rate for actor head
29 learning_rate_critic: 0.0003 # Learning rate for critic head
30 value_loss_coef: 0.5 # How much is the value loss weighted?
31 entropy_coef: 0.0 # How much is the entropy loss weighted?
32 max_grad_norm: 0.5 # Maximum norm for gradient clipping
33 network_size: 256 # Number of neurons in each hidden layer
34 normalize_rewards: 1 # Whether to normalize rewards or not (1 = True, 0 = False)
35 init_method: "orthogonal" # How will be the neural networks initialized?