Some MoGo-points-of-views around these 5 points:
Perhaps my understanding of Mogo from the thesis is incorrect. From a
certain standpoint it makes very limited usage of heuristics and seems
to rely solely several published details (in the thesis):
1) UCT+simulation
2) learned pattern weights via self-play using TD(lambda).
3) Proximity heuristics. (which is something I do not quite understand on a
deep level as to why it improves play).
4) RAVE knowledge recycling between trees.
5) The dragon heuristic.
1) does not hold in mogo anymore, because there's no upper-confidence
terms in MoGo anymore.
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).
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 :-)
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).
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).
Olivier
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
