Re: [Bug-gnubg] Intelligent Analysis

[Joern Thyssen]

On Tue, Sep 17, 2002 at 10:57:45AM -0300, Albert Silver wrote
> Hi,
>  
>             I have been thinking. Right now one of the great boons about
> the engines is how strong they are. Especially interesting is that they
> evolved their knowledge by themselves and that way were able go beyond
> us. The most frustrating part of it all is that they cannot share this
> with us in terms we can really understand, so we are forced to accept
> they are correct, and then try to explain it to ourselves. Fine, but it
> doesn't need to be that way for everything.

>             Take cube decisions. 
[..]
> game will be for the match. Your MWC should be X%"

This kind of information is available from Analyse->Market Window.

> The situations are more than a few, but there are a few that appear far
> more often than others, and those initially could be scripted. In order
> to not drown the Double analysis window, a button "Explanation" could be
> added should the player want it. In time, more cases would be covered. 
> Perhaps the program can never properly tell me that breaking my anchor
> is a catastrophe (though I have my thoughts on this too), but this is
> outside the domain of purely evaluation aspects (in other words its
> weights) and could and should become a part of a teaching tool IMO.
> Any thoughts? 

It's very hard to implement this. We would have to put logic into gnubg
to tell it what positional parameters to look for, since the neural net
is a "black-magic" evaluator. gnubg is able to give you the derivative of
the equity with respect to the pseudo-inputs, but I'm not sure that
information you can derive from that.

When humans decide what to move we often have to weigh contradictary
inputs, e.g., decide whether to run if we're leading the race but have
more men back, or deciding a risky hit when miles behind in the race and
with a weaker board than the opponent. The neural net evaluator weighs
these factors, but I don't think we can deduce what weights it used. We
can merely say that apparently it was more important to run than stay,
but we can't say why. I think that is one of the drawbacks of neural
bots versus older rule-based bots.

> BTW, I said I had my thoughts on positional explanations too. In most
> cases, it's true there is no way it can help, but sometimes one sees
> very clear things in the lists of moves and their evaluations. For
> example, I see that my "solid" move is a blunder of over 100 and placed
> 4th. Interestingly, every single one of the three first moves involves
> hitting a specific blot. It doesn't take Sherlock Holmes to conclude
> that NOT hitting is a colossal blunder. Or another slight variation, one
> sees that every move involving breaking the anchor leads to an enormous
> loss of equity. There are different reasonable candidate moves
> available, but breaking the anchor is not one of them. I don't know if
> it is possible, but even such light explanations could be of use to
> players at the lower ranks. If you don't scrutinize the lists of moves,
> one could easily miss these common points.

Yes, unfortunately you still have to do some work yourself :-)

Jørn

-- 
Joern Thyssen, PhD
Vendsysselgade 3, 3., DK-9000 Aalborg, Denmark
+45 9813 2791 (private) / +45 2818 0183 (mobile) / +45 9633 7036 (work)
Note: new mobile number!