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: "HalfCheetah-v5"
9name: "SAC"
10dir: "logs"
11notes: "Training HalfCheetah-v5 using SAC"
12monitor_gym: "False"
13config:
14 # Environment, logging and saving control
15 environment: "HalfCheetah-v5" # Environment to use
16 algorithm: "SAC" # What kind of algorithm to use?
17 save_dir: "models" # Where to save model?
18 save_name: "sac_half_cheetah" # 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: 1_000_000 # How many steps can fit into the memory?
23 learning_rate_q: 0.0003 # Learning rate for Q-Network
24 learning_rate_actor: 0.0003 # Learning rate for Actor network
25 tau: 0.005 # Interpolation factor in target network updates
26 warmup_steps: 5_000 # How many steps before agents starts optimising?
27 log_std_min: -5 # Std normalisation lower bound
28 log_std_max: 2 # Std normalisation upper bound
29 batch_size: 256 # How many steps are sampled from memory when optimising?
30 gamma: 0.99 # Discount factor
31 network_size: 256 # Number of neurons in each hidden layer
32 max_grad_norm: 1.0 # Gradient clipping constant to prevent grad explosion
33 reward_scale: 1.0 # Rewards need to be scaled for entropy temperature
34 policy_update_frequency: 2 # How often will the policy be updated?
35 normalize_rewards: 0 # Whether to normalize rewards or not (1 = True, 0 = False)