Dear all,
I am currently developing a reinforcement learning agent for the optimal control of energy systems using python. Please, how can I plot the graph of the selected agent's cumulative reward vs episodes number and penalty values vs episode number during or after training?
Also, how can I train different agents using the same environment, and plot their cumulative reward vs episodes number on the same figure?
I have seen the application in many research articles on DRL.
Any recommendation, links, or python code sample (preferably python) for the Implementation will be appreciated.
Best Regards,
Michael.
Alabi Tobi Michael
Running an RL agent in the environment for N times and calculating the average return from the N runs is a decent technique to assess it. It is typical practice to repeat the above evaluation step throughout your training process and graph the average return as training progresses.
In addition have a look here also: https://www.researchgate.net/publication/303331678_Dynamic_Frame_skip_Deep_Q_Network/figures?lo=1&utm_source=google&utm_medium=organic
@Qamar UI Islam. Please, I will appreciate it if i can see a sample code on how to graph the avarage returns as training progresses.
Thanks!
Qamar Ul Islam
Thank you very much for the reply.
Do you mean I will write the matplotlib code after the last command for the RL agent before running the entire code?
Thanks
Hi Qamar Ul Islam ,
what I usually do is using wandb to plot everything.
It is pretty simple to use and setup, and there are a lot of tutorials in their own website (wandb.ai).
At the end of each episode, when done is set to True, i just log the cumulative episode reward. Following the same logic, I can plot other useful system metrics. For the y-axis, you can leave it blank (it will use a "step" variable, which is increased every time you call wanbd.log with the arg commit=True), or just use another variable to keep track of the current episode number.