Final presentation and final thoughts
As of yesterday we presented our final results and submitted the report on the project we have been working in during this summer. It was a very fun day as we also got to see what other groups have done, and it feels nice to be able to relax a bit before the studies start again in the autumn. Let’s talk abut how the final weeks went and pick up the loose ends from the last blog post!
This project has been a lesson in how to readjust your project in line with new results, and from the last blog post we changed the goalpost quite a bit. As it turned out we could not extrapolate much between different parameter values for different matrices as the correlations simply were not present in the small data set we analysed. Because of this we turned our focus to instead finding optimal parameters for single matrices.
What we then found was three problems with the Bayesian sampling technique which we had chosen:
- The sampling of the real results where sometimes uneven in workload which gave raise to outliers with inaccurate reward values and a skewed model as a result.
- Because the algorithm always sampled the theoretical maxima of the number of points it tended to get stuck early in a local maxima, pointing toward a need of
- The function surface of the statistical model was sometimes poorly fitted for our needs as fitting one good reward value was less prioritized than fitting many clustered samples.
As the project had already undergone a couple of changes when this was discovered and since gathering data was such a long process, we could not look further into these issues more than pointing them out to give further research some thoughts to start with. While a more thoroughly positive result would have been a fun way to end the summer, we are still happy with what we managed to do given the time frame. For a more detailed summary of the final algorithms and its issues, please see the video linked above with our final presentation.
If you have followed the soHPC blogs during this summer, I hope you have had an as intersting journey as we have had and that you have learned a lot!