Event-Based Anomaly Detection for Searches for New Physics
Round 1
Reviewer 1 Report
It is a very interesting paper on the search for new physics.
The complexity of the argument is explained in detail and in a convincing way.
In my humble opinion, it can be accepted.
Some typo corrections and/or style suggestions are contained in the attached file.
Comments for author File: 
Comments.pdf
Author Response
Thanks a lot for the careful reading of the manuscript. We have included all the suggested changes.
regards, S.Chekanov (for the authors)
Reviewer 2 Report
I want to thank the authors for this nice work. Please, I have few questions"
Why the outhors use R=0.4 not 0.5 or 0.7?
Why you put "minimum transverse energy of all jets was
40 GeV in the pseudorapidity range of |η| < 2.5?
900K events will be enough for training the network? what will be happened id this number increase??
What will be happen if we change the event generator like EPOS or Herwig we will get the same results?
The abstract need to improve, please add some sentences.
Line 16:above-mentioned techniques => techniques as mentioned above
Line 16: are searches for => are searching for
line 36:In this paper we how a fraction of data => In this paper we discuss how a fraction of data
Line 41: from background => from the background
line 67:produce various types of particles =>produce various particles
Author Response
Thanks a lot for the careful reading of the manuscript. We have included your changes in the grammar. This manuscript also includes change from the 1st referee. Concerning your questions:
>Why the authors use R=0.4 not 0.5 or 0.7?
We added a text that tells that we use the definition adopted in ATLAS (in most of their papers).
>Why you put "minimum transverse energy of al jets was
>40 GeV in the pseudorapidity range of |η| < 2.5?
This was just our choice during the event generation. We do not think it should change any results. The |eta| is is quite standard cut for the LHC experiments.
> 900K events will be enough for training the network? what will be happened id this number increase??
Such variations in the numbers of events have been done in the past. There is no particular change when reducing the number of events. But we consider this 900k is quite large number that should be sufficient to pick the main features during ML learning.
> What will be happen if we change the event generator like EPOS or Herwig we will get the same results?
We did not investigate this. The purpose of this paper is to give an illustration of the method using available (to us) Monte Carlo, rather than a comprehensive studies that involve several MC types (Herwig/EPOS, presumably, may address systematic uncertainty issues for background estimates).
The changes discussed below are implemented. We only feel that the line 16 should stay as it is now, i.e. "are searches for " looks a better version for us.
>The abstract need to improve, please add some sentences.
The abstract was improved and extended.
>Line 16:above-mentioned techniques => techniques as mentioned above
>Line 16: are searches for => are searching for
>line 36:In this paper we how a fraction of data => In this paper we discuss how a fraction of data
>Line 41: from background => from the background
>line 67:produce various types of particles =>produce various particles
These comments were included.
Thanks a lot for the comments
S.Chekanov (for the authors)
