Thanks for the reply.
2) learned pattern weights are not learnt through
> TD(lambda). RLGO is not
> used in mogo. It was used a long time ago. Hand-designed
> heuristics are
> much more efficient (in particular after heavy
> cluster-based tuning of
> coefficients).
I am not entirely sure what you mean here by tuning coefficients do the
heuristics in question require some form of parameterization? How are these
parameters tuned?
>
> 3) holds in mogo, and in all efficient Monte-Carlo based
> techniques. This
> is particularly important, as seemingly knowing the last
> point if
> important for a correct evaluation of a position in the
> case of bounded
> computational resources. By the way, this might be true for
> humans, also.
> I'm afraid it's difficult to find sufficiently many
> human players and
> situations for organizing an experiment about that :-)
>
This seems to be the case and I still do not really on some level understand
why. Since with the chinese go rules the board should be effectively stateless
(exempting ko information) all the information be contained in the the current
position. Why additional information is needed i.e. an annotation of the last
played move is required on some level is a mystery to me.
> 4) The rave heuristic only "migrates"
> informations from one node to nodes
> "above" that node - never to brothers, cousins,
> sons, and so on. Many
> attempts have been done here and elsewhere around that
> without success.
>
> By the way, parallelizations (both multi-core and MPI) are
> *very*
> efficient. The use of huge clusters could probably give
> much better
> results than the current limit of mogo (between 1k and 1d
> on KGS with
> 64 quad-cores and myrinet switch).
This is obviously quite an impressive feat. However it is also on some level a
bit disappointing to me that it will be sometime before we see strong desktop
programs since the computational resources required is somewhat prohibitive.
>
> Another point which was not in the thesis of Sylvain and
> which is also
> central is the use of patterns in the tree. This is used
> since a long time
> in CrazyStone and Mango, we are currently developping this
> part in MoGo,
> and this is quite efficient (but involves tedious tuning of
> several
> parameters).
>
I am sure I understand the distinction here between patterns in the tree and
patterns used in the heavy playouts. I guess by this you mean they are not the
same thing.
I am in general basically curious how tuning of parameters occurs i.e. how you
fit the parameters in a given situation if things like reinforcement learning
are not used (which my understanding is is sort of an automated procedure for
fine tuning various parameters in the system).
Regards,
Carter.
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
