chenhongge / sa_dqn Goto Github PK

The paper mentions ‘We implement Double DQN [72] and Prioritized Experience Replay [58] on four Atari games’, but the actual code implementation does not implement Double DQN, but rather Dueling DQN. So I want to know if there was an error in the description in the paper.

I'm trying to evaluate the pre-trained RoadRunner model. The return I got from running
python test.py --config config/RoadRunner_cov.json test_config:load_model_path=models/RoadRunner-convex.model is around one thousand (as shown in the figure below), which is far from the reported 44638.0±7367.0.
I'm wondering what are the causes. Is there any updated environment or package that affects the performance?

Dear authors:
Thanks for sharing this code! This is a great work! However, when I try to train the robust agents with PGD solver, I cannot get the same level of agents of the paper.
After testing these agents, only get:

• 0.0 +- 0.0 average reward in the Bankheist environment.
  
• 0.0 +- 0.0 average reward in the RoadRunner environment.
  
• 21.46 +- 1.6150541786577934 average reward in the Freeway environment.
  
• -21.0 +- 0.0 average reward in the Pong environment.
  

Maybe my configured environment is the cause? Could you provide the exact version of python and main packages (torch, gym, numpy etc.)?
Or what else might be the cause？

Looking forward to hearing from you! Thank you!