This repo contains code accompaning the paper, Escaping from Zero Gradient: Revisiting Action-Constrained ReinforcementLearning via Frank-Wolfe Optimization (UAI 2021). It includes code for running the NFWPO algorithm presented in the paper, and other baseline methods such as DDPG+OptLayer, DDPG+Projection, DDPG+Reward Shaping, SAC+Projection, PPO+Projection, TRPO+Projection, FOCOPS.

This code requires the following:

• python 3.*
• TensorFlow v2.0+
• pytorch+cuda
• mujoco-py
• Gurobi
• CVXPY
• CvxpyLayer
• QPTH

To run the code, enter the directory of the corresponding environment, and run the following command: (Change ALGORITHM_NAME to the corresponding algorithm, which includes NFWPO, DDPG_Projection, DDPG_RewardShaping, DDPG_OptLayer, SAC_Projection)

python3 [ALGORITHM_NAME].py

(To run other baselines such as PPO+Projection, TRPO+Projection, and FOCOPS, please refer to the description below.)

Following are the examples for running the experiments in Ubuntu.

To run the experiments metioned in Secion 4.1, please follow the instructions below:

1. Enter the directory BSS-3:

cd BSS-3

2. Set the random seed arg_seed between 0-4 in Line 25 of NFWPO.py.
3. Use the following command to train NFWPO:

python3 NFWPO.py

4. To run other baseline methods, set the random seed arg_seed between 0-4 in DDPG_Projection.py, DDPG_RewardShaping.py, and run the corresponding command:

python3 DDPG_Projection.py

python3 DDPG_RewardShaping.py

5. The result is shown in Figure 1.

1. Enter the directory BSS-5:

cd BSS-5

2. Set the random seed arg_seed between 0-4 in NFWPO.py.
3. Use the following command to train NFWPO:

python3 NFWPO.py

4. To run other baseline methods, set the random seed arg_seed between 0-4 in DDPG_Projection.py, DDPG_RewardShaping.py, and set random_seed in DDPG_OptLayer. Then run the corresponding command:

python3 DDPG_Projection.py

python3 DDPG_RewardShaping.py

python3 DDPG_OptLayer.py

5. The result is shown in Figure 2.

1. Enter the directory NSFnet/src/gym:

cd NSFnet/src/gym

2. Set the random seed arg_seed between 0-4 in NFWPO.py.
3. Use the following command to train NFWPO:

python3 NFWPO.py

4. To run other baseline methods, set the random seed arg_seed between 0-4 in DDPG_Projection.py, DDPG_RewardShaping.py, and set random_seed in DDPG_OptLayer. Then run the corresponding command:

python3 DDPG_Projection.py

python3 DDPG_RewardShaping.py

python3 DDPG_OptLayer.py

5. The result is shown in Figure 3.

To run the experiments metioned in Secion 4.3, please first enter the directory Reacher for Reacher with nonlinear constraints, and enter Halfcheetah-State for Halfcheetah with state-dependent constraints.

cd Reacher

cd Halfcheetah-State

To run NFWPO, DDPG+Projection, DDPG+Reward Shaping, DDPG+OptLayer, please refer to the description in the previous experiment.

To run SAC+Projection, use the following command:

python3 SAC+Projection

To run TRPO+Projection, PPO+Projection:

# For Halfcheetah-state task
python3 PPO_TRPO_Projection/PPO_Projection_Halfcheetah_State_Relate_gym.py 
python3 PPO_TRPO_Projection/TRPO_Projection_Halfcheetah_State_Relate_gym.py

# For Reacher task
python3 PPO_TRPO_Projection/PPO_Projection_Reacher_State_Relate_gym.py 
python3 PPO_TRPO_Projection/TRPO_Projection_Reahcer_State_Relate_gym.py

To run FOCOPS:

# For Halfcheetah-state task
python3 FOCOPS/focops_main_cheetah.py

# For Reacher task
python3 python3 FOCOPS/focops_main_reacher.py

The result is shown in Figure 4-5.

To run the experiments metioned in Appendix D.3, please first enter the directory Halfcheetah-CAPG.

cd Halfcheetah-CAPG

Then run the following commands for corresponding baselines:

# CAPG+PPO
python3 CAPG_PPO_Halfcheetah_bound_constraints.py

# CAPG+TRPO
python3 CAPG_TRPO_Halfcheetah_bound_constraints.py

The result is shown in Figure 6.