PPO

Reproduce PPO with PARL

Based on PARL, the PPO algorithm of deep reinforcement learning has been reproduced, reaching the same level of indicators as the paper in Atari benchmarks.

Include following approach:

• Clipped Surrogate Objective
• Adaptive KL Penalty Coefficient

PPO in Proximal Policy Optimization Algorithms

Mujoco games introduction

Please see here to know more about Mujoco games.

Benchmark result

How to use

Dependencies:

• python3.5+
• paddlepaddle>=1.5.1
• parl
• gym
• tqdm
• mujoco-py>=1.50.1.0

Start Training:

# To train an agent for HalfCheetah-v2 game (default: CLIP loss)
python train.py

# To train for different game and different loss type
# python train.py --env [ENV_NAME] --loss_type [CLIP|KLPEN]