Reinforcement Learning with Learned Fourier Features
State-space Soft Actor-Critic Experiments
Move to the state-SAC-LFF repository.
cd state-SAC-LFF
To install the dependencies, use the provided environment.yml file
conda env create -f environment.yml
To run an experiment, the template for MLP and LFF experiments, respectively, are:
python main.py --policy PytorchSAC --env dm.quadruped.run --start_timesteps 5000 --hidden_dim 1024 --batch_size 1024 --n_hidden 3
python main.py --policy PytorchSAC --env dm.quadruped.run --start_timesteps 5000 --hidden_dim 1024 --batch_size 1024 --n_hidden 2 \
--network_class FourierMLP --sigma 0.001 --fourier_dim 1024 --train_B --concatenate_fourier
The only thing that changes between the baseline is the number of hidden layers (we reduce by 1 to keep parameter count roughly the same), the network_class, the fourier_dim, sigma, train_B, and concatenate_fourier.
Image-space Soft Actor-Critic Experiments
Move to the image-SAC-LFF repository.
cd image-SAC-LFF
Install RAD dependencies:
conda env create -f conda_env.yml
To run an experiment, the template for CNN and CNN+LFF experiments, respectively, are:
python train.py --domain_name hopper --task_name hop --encoder_type fourier_pixel --action_repeat 4 \
--num_eval_episodes 10 \--pre_transform_image_size 100 --image_size 84 --agent rad_sac \
--frame_stack 3 --data_augs crop --critic_lr 1e-3 --actor_lr 1e-3 --eval_freq 10000 --batch_size 128 \
--num_train_steps 1000000 --fourier_dim 128 --sigma 0.1 --train_B --concatenate_fourier
python train.py --domain_name hopper --task_name hop --encoder_type fair_pixel --action_repeat 4 \
--num_eval_episodes 10 \--pre_transform_image_size 100 --image_size 84 --agent rad_sac \
--frame_stack 3 --data_augs crop --critic_lr 1e-3 --actor_lr 1e-3 --eval_freq 10000 --batch_size 128 \
--num_train_steps 1000000
Proximal Policy Optimization Experiments
Move to the state-PPO-LFF repository.
cd pytorch-a2c-ppo-acktr-gail
Install PPO dependencies:
conda env create -f environment.yml
To run an experiment, the template for MLP and LFF experiments, respectively, are:
python main.py --env-name Hopper-v2 --algo ppo --use-gae --log-interval 1 --num-steps 2048 --num-processes 1 \
--lr 3e-4 --entropy-coef 0 --value-loss-coef 0.5 --ppo-epoch 10 --num-mini-batch 32 --gamma 0.99 \
--gae-lambda 0.95 --num-env-steps 1000000 --use-linear-lr-decay --use-proper-time-limits \
--hidden_dim 256 --network_class MLP --n_hidden 2 --seed 10
python main.py --env-name Hopper-v2 --algo ppo --use-gae --log-interval 1 --num-steps 2048 --num-processes 1 \
--lr 3e-4 --entropy-coef 0 --value-loss-coef 0.5 --ppo-epoch 10 --num-mini-batch 32 --gamma 0.99 \
--gae-lambda 0.95 --num-env-steps 1000000 --use-linear-lr-decay --use-proper-time-limits \
--hidden_dim 256 --network_class FourierMLP --n_hidden 2 --sigma 0.01 --fourier_dim 64 \
--concatenate_fourier --train_B --seed 10
Acknowledgements
We built the state-based SAC codebase off the TD3 repo by Fujimoto et al. We especially appreciated its lightweight bare-bones training loop. For the state-based SAC algorithm implementation and hyperparameters, we used this PyTorch SAC repo by Yarats and Kostrikov. For the SAC+RAD image-based experiments, we used the authors' implementation. Finally, we built off this PPO codebase by Ilya Kostrikov.
64 Dec 27, 2022
4 Sep 24, 2022
72 Dec 05, 2022
205 Jan 09, 2023
194 Dec 22, 2022
355 Jan 06, 2023
42 Jul 25, 2022
6 Feb 12, 2022
103 Dec 20, 2022
829 Dec 31, 2022
54 Nov 21, 2022
87 Nov 30, 2022
0 Dec 06, 2021
486 Dec 20, 2022
111 Jan 05, 2023
13 Jul 26, 2022
393 Dec 29, 2022
1.9k Dec 15, 2022
39 Dec 27, 2022
774 Dec 31, 2022