04:20 < luffy1996> @Zoq , I had one question. This will also give me more idea how to proceed for my previous PR on policy gradient methods.

04:24 < luffy1996> One of the most important thing is gradient ascent for policy networks. How should I proceed with this. I do have one idea. I think making a gradient ascent with the help of https://github.com/mlpack/mlpack/blob/master/src/mlpack/core/optimizers/sgd/update_policies/vanilla_update.hpp is possible

04:24 < luffy1996> What I think is , doing

04:25 < luffy1996> iterate += stepSize * gradient; // line 58

04:25 < luffy1996> @rcurtin, @zoq do give me your idea about this

09:55 < zoq> luffy1996: Agreed, that is probably the best solution.

10:45 < luffy1996> Should I go ahead and add it in the optimzers . I think it is not available with mlpack

10:49 < zoq> luffy1996: Yes, please feel free, not sure sgd is the best folder for that, maybe we should create a new folder sga?

10:59 < luffy1996> I was having the same doubt. I am going with making a new folder as gradient_ascent .

11:00 < luffy1996> @zoq, You would like to add the recent PR I have submiiterhttps://github.com/mlpack/mlpack/pull/1361

11:33 < zoq> luffy1996: Will, look at it once I get a chance.

14:12 < robertohueso> When I try to bound the error of an algorithm that uses a tree, I do it using MaxDistance and MinDisntance methods but when call Center methods, that point's distance exceeds the error, does this make sense?

14:12 < rcurtin> robertohueso: can you explain what you mean a little further?

14:14 < robertohueso> In my particular case MaxDistnace = MinDistance (I guess this means the node has just 1 point) but Center is different from that point

14:14 < robertohueso> rcurtin: I'm particularly working on your answer to my PR #1301

14:15 < rcurtin> right, I know the context for KDE

14:15 < rcurtin> what is the tree type?

14:15 < robertohueso> kd-tree

14:16 < rcurtin> hmm, now that is interesting and unexpected, but I guess it is possible

14:16 < rcurtin> if a kd-tree node has only one point, then the size of the bound should be only that single point

14:16 < rcurtin> what is FurthestDescendantDistance() for that node? it should be 0.0

14:17 < rcurtin> also, is it possible that MaxDistance == MinDistance and both those are some invalid distance like NaN or inf? (I guess in that case the == operator would not be true)

14:17 < rcurtin> or are they both 0?

14:24 < robertohueso> FurthestDescendantDistance = 0.0

14:24 < robertohueso> MaxDistance = MinDistance = 3.100000000000000088817841970012523233891 so it seems to be valid

14:26 < robertohueso> Maybe I'm dumb but I can get the point on that node by using querySet.col(queryNode.Descendant(0)) where querySet is an arma::mat that holds query points right?

14:31 < rcurtin> yeah, that should work just fine

14:32 < rcurtin> but to me it is strange that Center() would return something different

14:32 < rcurtin> you could also check and see if the HRectBound has width 0 in each dimension

14:32 < rcurtin> I guess it must not, since Center() is returning something else

14:47 < robertohueso> HRectBound seems to have width 0

15:05 < rcurtin> that's strange, can you check to see if the point Descendant(0) even falls within the HRectBound?

15:05 < rcurtin> I think you can use .Contains() for this

15:11 < robertohueso> I just checked, it's NOT contained

15:15 < rcurtin> now that is strange, I think there may be some bug here

15:15 < rcurtin> this is surprising to me, because the trees are pretty rigorously tested

15:22 < robertohueso> I'm going to check again my code, it seems more likely that I made some mistake

15:25 < rcurtin> sounds good, but remember it is always possible there is a bug in the tree code too, so don't discount that possibility in your searching :)