Hyperparameter ballpark for symbolic envs
jendelel opened this issue · comments
Hi,
Thanks for porting PlaNet to PyTorch. Great work! I would like to use PlaNet for planning in maze-like environments. However, I can't get it to work reasonably even on classic gym Mujoco tasks such as Ant-v2. Have you tried running on symbolic environments? There are many hyperparameters and I don't know where to start tuning.
This is what I tried:
Options
id: Ant
seed: 1
disable_cuda: False
env: Ant-v2
symbolic_env: True
max_episode_length: 1000
experience_size: 1000000
activation_function: relu
embedding_size: 64
hidden_size: 64
belief_size: 4
state_size: 30
action_repeat: 1
action_noise: 0.3
episodes: 1000
seed_episodes: 5
collect_interval: 20
batch_size: 50
chunk_size: 50
overshooting_distance: 50
overshooting_kl_beta: 0
overshooting_reward_scale: 0
global_kl_beta: 0
free_nats: 3
bit_depth: 5
learning_rate: 0.001
learning_rate_schedule: 0
adam_epsilon: 0.0001
grad_clip_norm: 1000
planning_horizon: 12
optimisation_iters: 10
candidates: 1000
top_candidates: 100
test: False
test_interval: 25
test_episodes: 10
checkpoint_interval: 50
checkpoint_experience: False
models:
experience_replay:
render: False
Thanks in advance for you help,
Lukas
Sorry, I've not really tried to run PlaNet on other domains, so don't have any ideas. Since this is a port, you should be able to open an issue on the original repository and see if the original author can provide some guidance.
@jendelel , I also try to use PlaNet/Dreamer to run symbolic envs like Ant-v2 and HalfCheetah-v2. I would like to ask if you have found suitable hyperparameters? Thanks a lot!
Hi, not really! I moved away from Planet back then. If you're working with it now, I'd highly recommend using DreamerVX over Planet. It's a series of follow up papers.