TalkChess.com

Thanks for the link!

Neural Networks are black boxes. That's why you should not use them.

cdani wrote:https://www.technologyreview.com/s/6040 ... art-of-ai/

This has been an issue from the very beginning with respect to artificial neural networks. They find patterns in training data but the programmer doesn't necessarily have any idea what the salient features of the training data are (or the useless features). This is less true of small networks but large, sophisticated NNs can be intractable for the programmer.

I still have a working copy of BrainMaker 2.0 for DOS from 1989.

https://calsci.com/BrainIndex.html

Except for me! (But I don't know why)

As Arthur C. Clarke said, "Any sufficiently advanced technology is indistinguishable from magic."

--Jon

Yeah.. this is a bit counter-intuituve. One would say "the programmer knows it all", but the concept of training set makes stuff quite murky, as the algorithm calibrates in its own way.

Still, it also happens in heuristic chess engines. Do we really know why they choose this or that move?

whereagles wrote: Still, it also happens in heuristic chess engines. Do we really know why they choose this or that move?

Only to a point, as most changes we try don't work.

cdani wrote:

whereagles wrote: Still, it also happens in heuristic chess engines. Do we really know why they choose this or that move?

Only to a point, as most changes we try don't work.

Says the guy with a 3200ELO monster as his creation.

whereagles wrote:Yeah.. this is a bit counter-intuituve. One would say "the programmer knows it all", but the concept of training set makes stuff quite murky, as the algorithm calibrates in its own way.

Still, it also happens in heuristic chess engines. Do we really know why they choose this or that move?

Indeed. With feedback in a complex system you often get a chaotic system where the programmer knows for example the formula of a fractal, or a limited set of laws of physics but the results are a practically infinite complexity fractal, or universe.
Strangely not so easy to apply to chess. Because of it's tactical nature? Though it should only be a matter of time before pattern recognition becomes useful in computerchess now. I fool around with it from time to time where the idea ofcourse is for it to recognise tactical or big winning/losing positional themes before they are searched which can then be applied to an evaluation function. The problem is to search more than 3 nodes per second and I'm new to pattern recognition so I have no idea what I'm doing. (Read: it doesn't recognise anything.)

On a personal note I must have gotten my first book on fractals when I was 11. (For Dutch readers: Vuiks verhandelingen.) Thought I had gotten the anwsers to the universe and everything. If we knew this why wasn't it on tv and on the news every night?