Thread history

From Talk:Firestarter/Version History
Viewing a history listing
Jump to navigation Jump to search
Time User Activity Comment
14:15, 27 October 2017 Cb (talk | contribs) New reply created (Reply to scale standard deviations to 1)
10:56, 27 October 2017 Xor (talk | contribs) New reply created (Reply to scale standard deviations to 1)
10:26, 27 October 2017 Cb (talk | contribs) New reply created (Reply to scale standard deviations to 1)
01:52, 27 October 2017 Xor (talk | contribs) New thread created  

scale standard deviations to 1

Good to hear that some experiment was done with this rule-of-thumb. Anyway, did you find that improve your performance? Or some weight re-tune is needed imo.

Btw, are you using some online standard deviations algorithm and update the knn weights on the fly, or just do it offline with pre-collected data?

Xor (talk)01:52, 27 October 2017

Yes, I think you mentioned this rule-of-thumb somewhere on the wiki, (or was it somebody else? not sure now) thanks for that. I found that improve my performance. However, I focused on becoming stronger against surfers, which made me weaker against non-surfers.

I just run a battle against some bot, collected the standard deviations and afterwards divided each attribute by the respective std deviation. I found that now the standard deviations remain pretty close to 1 against most bots.

Cb (talk)10:26, 27 October 2017

Well, that's the easiest way I can think of... And it really surprised me that the standard deviation is almost the same for most bots. If so, then I think normalising standard deviation may not be that useful — since we already tune weights very hard. But for neural networks, I think this is invaluable, since we don't tune weights manually, and that would affect the learning speed a lot.

Xor (talk)10:56, 27 October 2017

I never thought about normalizing standard deviation for neural networks, because they should adjust themselves to the attributes anyways. But yeah, the learning speed is probably affected by it, I should consider that for further NN experiments :D

Cb (talk)14:15, 27 October 2017