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: "Hopper-v5"
9name: "TD3"
10dir: "logs"
11notes: "Training Hopper-v5 using TD3"
12monitor_gym: "False"
13config:
14 # Environment, logging and saving control
15 environment: "Hopper-v5" # Environment to use
16 algorithm: "TD3" # What kind of algorithm to use?
17 save_dir: "models" # Where to save model?
18 save_name: "td3_hopper" # 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 memory_size: 1000000 # How many steps can fit into the memory?
23 learning_rate_q: 0.0001 # Learning rate for Q-Network
24 learning_rate_actor: 0.0001 # Learning rate for Actor network
25 tau: 0.005 # Interpolation factor in target network updates
26 warmup_steps: 10_000 # How many steps before agents starts optimising?
27 batch_size: 256 # How many steps are sampled from memory when optimising?
28 gamma: 0.99 # Discount factor
29 exploration_noise: 0.1 # Noise that is added to actor during action selection
30 policy_noise: 0.2 # Noise that is added to actor during network optimization
31 noise_clip: 0.5 # Min/Max noise value
32 policy_interval: 2 # How often will the actor and target networks be updated?
33 network_size: 256 # Number of neurons in each hidden layer
34 max_grad_norm: 0.5 # Gradient clipping constant to prevent grad explosion
35 reward_scale: 1.0 # Used for reward scaling
36 normalize_rewards: 0 # Whether to normalize rewards or not (1 = True, 0 = False)