search

OLA2022-Group-12 / OLA2022-Project

Project for the Online Learning Applications class at Politecnico di Milano a.a. 2021-2022
2 stars 0 forks source link

Alphaless 2.0 (the real Alphaless learner) #24

Closed raul-singh closed 2 years ago

raul-singh commented 2 years ago

Here is the alphaless learner. It looks like it works correctly. The Alphaless performs slightly better than the AlphaUnitsless in our example environment. If we use a more unbalanced environment the performance gap increases, as it should be. They are both implemented in GPTS and GPUCB1 and as to be expected on average the TS one performs slightly better. I did extensive hyperparameter tuning to get this performance but I think we can still improve it. Somehow the average performance of the alphaless today dropped a bit and I don't quite understand why. But everything is working correctly anyway. Even after the merge I'll continue doing some hyperparams tuning. One thing we could do in the future is maybe use always the same graph for our tests since right now a new graph is generated every single time, and this can lead to inconsistencies (and maybe it's the cause of that slight performance drop I mentioned)

raul-singh commented 2 years ago

@davide-rigamonti-polimi I'm sorry to bother, but without merging this branch I cannot continue the work on the environment and since pretty much everyone else need the my work to continue their assignment I don't think it's a good idea to stall for this long without any activity here. We should come to a concusion.

daviderigamonti commented 2 years ago

I was waiting for @barskern to conclude the previous thread by sharing his opinion.

However, if you prefer to merge right away let's do it.