G’day! The past two months have seen the first bit of MVA application in my research. I’ve also had the pleasure of helping to supervise three students from IST, the local university, who’ve been working on summer internships at LIP. The students (António, João, and Ricardo) have expressed interest in detailing their work as guest-posts on here, so I’ll steer clear of a full description of the project, and simply give an overview:
During the first few months at LIP I had produced some basic MC samples and developed some code in Root to analyse them and save events which satisfied the selection criteria. Then I had attended various schools to build up my knowledge of MVAs and statistics. Lately I had been on secondment in Oxford, and helped to produce some more advanced MC samples.
The internship project aimed to reconcile these three areas of work, by running my final-state selection code over the new MC samples, and then applying MVAs to improve event classification and to develop a regressor for the di-Higgs mass.
Inspired by the seminars from the recent school in Lund, I decided to use Python implementations for the MVAs (mostly SK-Learn and Keras), and developed a simple BDT in a Jupyter Notebook for the students to use as a starting point. Their task was to improve upon it by adding new variables, preprocessing the data, optimising MVA hyper-parameters, switching to neural-networks, and any other methods they could come up with.
Another approach was to try and avoid the need for the final-state selection, which currently rejects a lot of signal events (about 99/100 events), by trying to ‘go full MVA’. My idea was to feed in generic information from the signal and background sample (4-vectors of the five hardest jets and whether they were b- or tau-tagged, the three hardest leptons, the hardest photon, HT variables, etc.). Pleasingly, the MVAs were able to offer event discrimination, and the results were promising. Hopefully more on this later on.
I’ve been really pleased with how hard the students worked and the independence they showed in finding techniques and software to improve their MVAs. It’s also been a great learning experience for me, since it was the first time I’d been in a position of academic responsibility.
In other news, I’m off to Padova for a two-month secondment! I’ll be working with fellow AMVA4NP researchers Cecillia, Greg, and Pablo to begin pinning down the exact implementations of the classification and regression algorithms in preparation for the fast-approaching activity report.
Feature image taken from http://lazyprogrammer.me/.