https://arxiv.org/abs/1710.02896

1. Environment: Miniconda is recommended as pybox does not support pip

• python 2.7: print format might become an issue with python 3 but other than that, is fine
• numpy, scipy, matplotlib: up-to-date
• tensorflow 1.2 : higher versions are fine and TF-GPU compatible
• OpenAI gym and pybox: for gym, download the files in 'gym-files.tar.gz' and replace 'bipedal_walk.py(many other versions are provided in the tar file)' and 'time_limit.py' into the original files

2. Run default model (Our RDPG)

• learn and run: run 'gym_ddpg.py' - be sure to make proper 'checkpoint' files for both 'saved_' folders and 'gym_ddpg' folder inside 'results' directory
• record: run 'tester_r.py'
• display: run 'display.py' in 'results' directory

3. Other models

• DDPG(Feedforward network-based DPG): d3_9
• RDPG with parameter noise: r17_41_opt0
• Our RDPG with experience injection: TBA