[computer-go] Explanation to MoGo paper wanted.

[Richard Brown]

Gunnar Farnebäck wrote:
> Don wrote:
>> Of course now we just had to go and spoil it all by imposing domain
>> specific rules.  I have done the same and I admit it.    It would be fun
>> to see how far we could go if domain specific knowledge was forbidden as
>> an experiment.   Once patterns are introduced along with other direct Go
>> knowledge, it's still fun but it feels a bit "wrong", kind of like
>> cheating.   It's clear that when we do this, we introduce strengths and
>> weaknesses to the program,  making it a bit more fragile, less
>> "universal" or robust.  Stronger too, but more susceptible to
>> in-transitivity.
>
> I'm on the other side of this issue. In my opinion all kinds of go
> knowledge are fair game and I'm rather disappointed that so small
> amounts of domain specific knowledge have been merged with the UCT
> search approaches.

Thank you for articulating this viewpoint, Gunnar.

A program trained upon millions of patterns harvested from professional games
[appearing soon on a server near you, for some appropriate value of "soon"]
has the property that it _severely_ prunes the game tree according to what
might be described as "heuristics", but which heuristics are _learned_ by
several hundred feedforward networks.  [The heuristics are not hand-coded.]

The patterns comprise observations of what pros do, and more importantly,
what they don't do.  Selecting which features of go-behavior to measure,
and upon what scale to express those measurements, are both, of course,
crucial.  [The feature-extraction phase of pattern-recognition is a tricky
problem, regardless of domain:  Eschewing noise in favor of information.]

Once the feedforward networks have done their learning, the patterns themselves
are discarded.  [Thus there is no pattern "library", only the trained networks.]

Applying a UCT-layer atop such a severely-pruned game-tree seems promising.

I would re-phrase what Gunnar said:  I am optimistic that when UCT is
layered atop a strong program that has already pruned -- via learned,
domain-specific knowledge -- most of the fruitless branches from the
tree, such a program will become even stronger by virtue of the UCT-layer.

Don is right, too, in the sense that such a program is somehow less "pure"
than one which has obtained all its knowledge from first principles.

Nonetheless, a program that could not only play a decent game of go, but
somehow emulate the _style_ of a given professional would be of interest,
would it not?

Of even more interest, perhaps, would be a program that, given a game of
go between two (unidentified) professionals whose games it had studied,
whose styles it had _learned_, could then tell you who the two players
were, _even_if_it_had_never_before_seen_that_particular_game_.

Alternatively, if the identity of the players were known, such a program
could be expected to perform well (much better than random, anyway) at a
"guess-the-next-move" task.

Could a pure-UCT program ever perform in these ways?  Even if the pure-UCT
program were able to reliably and repeatedly defeat the domain-specific
program at playing go, it would not be able to perform the other tasks.

So, I must agree with Gunnar that the domain-specific approach is "fair game".

It all depends on which problem one is trying to solve, to begin with.

-- 
Rich