Plotted Reward Scale

[OscarHuangWind]

Hi, I am oscar and I do appreciate those source codes with integrating various algorithms.

I have tried to run the nature DQN with default setting through Pong and BeamRider environment and found that the reward scale is not as large as the one posted in main page.
For Pong Environment,
I just manually set the clip_rewards = False and got the final mean around 27.430 which is far from the max level(around 300) posted.

Is it due to difference hyper-parameters setting or may be due to some plotting techniques?
BTW, I will really appreciate if you can update the plotting code, Thank you!

[TianhongDai]

@OscarHuangWind Hi Oscar - Thank you for pointing out this problem. This is because of the setting of the environment:episode_life. The reward displayed in the terminal is not the real accumulated reward, the real accumulated reward is recorded in the logs/<envs> folder, and you need to use those files to plot the learning curve. Could you please check the file in that folder firstly?

In addition - I will upload / update the plot script. If you have further questions, please let me know.

[OscarHuangWind]

Hi, thank you for your quick reply.
I plotted .csv file in "logs/envs" and found that the reward scale is right. But still I don't understand the reason.

Currently I am doing research in DQN algorithms with multiple atari environments and suffered from the reward scale problem as well. Would you mind to write me an email and discuss more specifically? Or I will appreciate if you leave me an email address.

My email address is: [email protected]
Thank you in advance!

[TianhongDai]

@OscarHuangWind Hi Oscar, No Problem, I will drop you an email today.