Trouble is that it is very difficult to put certain concepts into mathematics. For instance: “well, I tried to find parameters that did a better job of minimizing that error function, but eventually I lost patience.” :-)
Neural network parameters are not directly humanly understandable. They just happen to minimize an error function on a sample of training cases that might not even be representative. So you want to reason “around” the NN by interrogating it in some way, and trying to explain the results. If anyone wants to pursue this research, I suggest several avenues. First, you could differentiate the output with respect to each input to determine the aspects of the position that weigh on the result most heavily. Then, assuming that you can compare the scale of the inputs in some way, and assuming that the inputs are something that is understandable in the problem domain, maybe you can construct an explanation. Second, you could construct a set of hypothetical different similar positions, and see how those results differ. E.g., make a set of examples by adding a black stone and a white stone to each empty point on the board, or removing each existing stone from the board, and then evaluate the NN on those cases, then do decision-tree induction to discover patterns. Third, in theory decision trees are just as powerful as NN (in that both are asymptotically optimal learning systems), and it happens that decision trees provide humanly understandable explanations for reasoning. So maybe you can replace the NN with DT and have equally impressive performance, and pick up human understandability as a side-effect. Actually, if anyone is interested in making computer go programs that do not require GPUs and super-computers, then looking into DTs is advisable. Best, Brian From: Computer-go [mailto:computer-go-boun...@computer-go.org] On Behalf Of Jim O'Flaherty Sent: Wednesday, March 30, 2016 4:24 PM To: computer-go@computer-go.org Subject: Re: [Computer-go] new challenge for Go programmers I agree, "cannot" is too strong. But, values close enough to "extremely difficult as to be unlikely" is why I used it. On Mar 30, 2016 11:12 AM, "Robert Jasiek" <jas...@snafu.de <mailto:jas...@snafu.de> > wrote: On 30.03.2016 16:58, Jim O'Flaherty wrote: My own study says that we cannot top down include "English explanations" of how the ANNs (Artificial Neural Networks, of which DCNN is just one type) arrive a conclusions. "cannot" is a strong word. I would use it only if it were proven mathematically. -- robert jasiek _______________________________________________ Computer-go mailing list Computer-go@computer-go.org <mailto:Computer-go@computer-go.org> http://computer-go.org/mailman/listinfo/computer-go
_______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
