I enjoyed reading this paper! It is a nice piece of work on a very interesting topic.

My main comments are related to i) comparison of algorithmic parameters and ii) extensions to 3D. Comment i) requires some work, but I think it should be relatively fast to do in a revision. Comment ii) can be discussed some more and left for future work.

i) Your paper contains a number of cases, but there are limited comparison of the suggested method using different tuning parameters m, k, alpha. I am guessing by tuning some of these one could have a greedy approach at one end versus a very deep one which is more time consuming at the other end. But I don't see much comparison of using various of these (extreme) inputs as it is now. I am also not sure how easy it would be to compare the suggested approach with ones like Q-learning or other RL / value iteration methods for your case?

ii) In practical mining operations, wouldn't there ordinarily be sequential 3D boreholes where one can choose and modify the drilling order / locations? One could also potentially stop data collection (and drilling) in one borehole after a certain depth (before the initial planned depth is reached). Along boreholes one could also have different data collecting frequency. The suggested strategy for collecting data seems a bit restricting in this setting - as it is 2D only in this paper. What more is needed or possible in 3D? 

iii) Some detail comments:

- Mark a_1 and a_2 on first axis of Figure 1, as well as have 'x' or similar as the axis label.

- l175: This is accounts?

-Around Table 1, I don't think all these comparison of AI and geo terminology are needed.

-In Sect 4.1 there is a discussion of "actions", and you state 'the agent may acquire measurements (data)'. But at this point in the presentation there is no observation terminology 'o'. Aren't the action here to mine or abandon?

-Not sure L(o_{t+1}|...) is defnied in l 225 expression (it comes much later, I think)?

-Algorithm 1, data line should have d <- d + e, e \sim N(0,\sigma_n^2)

-Sunberg and Kochenderfer, 2018 paper is not on the reference list?

- \sigma means several different things in the paper and can be a bit confusing.

-You often say Figure X below. You don't need the 'below' here.

-Would it be possible to color-code the histograms in Fig. 15 (+ similar ones) according to 'mine' or 'abandon' ? Couldn't you also have one bar for each outcome here, rather than bins 0-5, 5-10, etc.?

Thanks for a nice paper!

Dear authors,

Thanks for your interesting contribution. The manuscript is well written and very pleasant to read. The objectives are very clear and the method is rather well explained. I have a few suggestions and questions to clarify some points and facilitate the reproducibility of the work.

Line 56: could you explain what a non-sequential scheme could be in the context of mineral exploration, as it seems to contradict the previous sentence on line 38. It becomes clearer though, when reading the following paragraphs.

Last paragraph of section 1:  Which of the mentioned approaches did you select for your demonstration?

Section 4.2: how is the state space initialized?

Line 289: should it be r(s,a)= -Cost(s,a) or Cost(s,a)=-Cmeasurements to be consistent with the substraction of Cextraction in the profit?

Table 2: where does d come from?  Formatting: should it be an algorithm rather than a table object? See e.g. the example in the latex template

Line 356: ‘At each time step’

Line 357: ‘The full tree is constructed’ – in the case of the POMDP ?

Line 359: by trial trajectory, do you mean a branch of the tree or realization of the full tree? How is the (partial) tree generated? What is the prior over the trajectory length , and between the different actions (explore further, mine or abandon)

Line 378: previous visits cumulated over the previous iterations t ?

Lines 423 to 425 and figure 9 bottom right panel: can you clarify the stooping criteria as at =5 the mean of the ensemble decreases and is getting smaller than the extraction cost. Can you also clarify how the value of gained information is assessed?

Figure 13: missing scale for the mean average error