Trust Region Proximal Policy Optimization with Smooth Roll-Back

This is the implementation of paper Trust Region Proximal Policy Optimization with Smooth Roll-Back (TR-PPO-SRB) algorithm on Pytorch version, the corresponding paper is submitted to IEEE. See the record in https://www.researchgate.net/publication/370809773_Trust_Region_Proximal_Policy_Optimization_with_Smooth_Roll-Back

Algorithms

• PPO-Clip
• PPO-KL
• Trust Region PPO (TR-PPO)
• PPO Smooth (PPO-S)
• Trust Region PPO with Roll-Back (TR-PPO-RB)
• Trust Region PPO with Smooth Roll-Back (TR-PPO-SRB)

Packages recommodation

• python=3.8.5
• torch>=1.13.1
• gym==0.10.5

Train agent with specific algorithm

For example, we train the agent with TR-PPO-SRB algorithm on the GPU with 8 parallel worker:

python main.py -GPU=True -number-of-parallel=8 -algorithm=TR-PPO-SRB

Test agent with specific algorithm

For example, we test the agent with TR-PPO-SRB algorithm:

python main.py -evaluate=True -algorithm=TR-PPO-SRB

Plot the training results

python main.py -plot=True -env-name=Swimmer-v2