Closed raul-singh closed 1 year ago
Btw, how is your work on sliding-window algorithm and on change-detection algorithm is going, @barskern ? I think we should wrap up this part on dynamic environment as soon as possible and focus on finishing touches and the presentation.
Somewhat slow, but with this PR merged I will have a great way of setting up some testing facilities which will make it a bit easier. Thank you!
Ok, everything should be solved! Let me know, if the part you were doing is problematic for you I could take the job and do it, since I have worked already a lot with the learners and have a good idea on how to implement sliding window and change detection.
I tried many ideas to implement abrupt changes in our environment but in the end the one that worked best is just a simple 5-line function. For the project's purpose we just need a simple abrupt change, nothing fancy. Therefore I implented a simple function that basically makes all the classes lose interest in a certain product.
As we can see here at day 25 an abrupt change kicks in, making product 1 very uninteresting to classes. This makes our algorithm perform much worse since it learned to put a lot of budget into subcampaign 1, being it very profitable. This is the result we were expecting. Obviously a dynamic algorithm should deal better with this abrupt change.
I also made a notebook (which is basically a copy of the LearnerPlayground notebook) to try out the dynamic environment.
Btw, how is your work on sliding-window algorithm and on change-detection algorithm is going, @barskern ? I think we should wrap up this part on dynamic environment as soon as possible and focus on finishing touches and the presentation.