A Penetration Method for UAV Based on Distributed Reinforcement Learning and Demonstrations
Round 1
Reviewer 1 Report
Concept issues:
1) There are two parameters that can be used as command outputs of the Reinforcement learning algorithm: overload and attitude angle. You choose the overload. Why? What are the advantages or drawbacks of using one or the other?
2)In lines 127-130 you say that the reinforcement leraning network need the position and speed information of both the UAV and the interceptor. This is OK in a simulation scenario, but is it realistic in a real scenario?
3) what is an antipersonnel warhead? why 400 m distance is safe? are not different types of interceptors?
4) in lines 214-216 you talk about efficiency. It is clear that using the pre-training is more efficient that not using it. However you no not present any information or results regarding the efficiency of the proposed algorithms.
5) Regarding the results shown in figure 7, you performed 500 experiments in total or 500 experiments of each case (pre-training; A3C; Pre-training-A3C; CPL) ? it is not clear
Technical issues:
1) The abstract do not reflect the contents and ideas of the paper and its structure and English problems result in a bad comprehension. Abstract should be rewritten.
2) Introduction (section 1) is really short and it only has some related work comments. I think the introduction needs to explain the problem, the scenario, etc. plus the related work and the following structure of the paper.
3) line 260 has an error (reference not found)
4) line 359: "Fix" should be "Fixed" ?
5) check for English errors and typos.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Why there is no comparison the other algorithms
The equations 2 and 4 are clear, I suggest to the authors to explain more
Why the authors do not specifying the system used for the positioning of UAV
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
