Valkyria9 bot

Time to congratulate Valkyria9 bot

http://www.littlegolem.net/jsp/message/message.jsp?msgid=1740488

Took me by surprise :)

As I read it, this was played by Magnus himself,

not by his bot (valkyria is a Go bot).

Of course we'd welcome another Havannah bot!

Should be an easy exercise for Magnus - since all the MC/UCT

machinery is already in place ... Of course I'm joking.

Correct link: christian was refering to one of these two games: game 1249149 and game 1246897 .

Of course I could write a Havannah program as well. It is tempting but I have no time to do that. I repeat: it will not happen!

But I am constantly thinking about how to make it when I play Havannah...

Hypothetically, I am not sure I could make it strong as Valkyria on Go. Also I think whatever I myself do then I play Havannah is not so super complicated. It just have to translated into the playouts. Uniformly random playouts I think is really ineffective in Havannah which a lot of long forced sequences that should be read out by the program.

forced sequences: indeed, but literally forcing (threat to win)

happens only at the very end (which is early on small boards only),

or with local ring threats.

I think the game is decided by local fights for group status

(can reach this side, can reach that corner) but these are

not absolute threats (you may give up one goal, to reach another)

so it requires some second-level (global) control.

Currently R-o-F does not implement this (guess what - I have no time).

Best - Johannes.

Yes, I was unclear with "forced sequences" in Valkyria for Go in the playouts all sorts of local "forced moves" are played deterministically. But they are not necessarily forced given the whole board. It is the tree part of the search that determines if those local patterns are Sente or not. In many ways I think there are similarities between go and Havannah.

So playouts should try out any reasonable local patterns. Threaten to make frameworks and prevent them. Tree search is what you call global control. It is not explicitly programmed but comes from the inherent power of monte carlo evaluation. A more explicit control can also be introduced by biasing tree search using more complex algorithms than the playouts do.

I feel some relief, in any case, that Valkyria wasn't a bot :)

:o)

you felt a bit scared? :P

I've lost base-8 and base-10 against Castro and Deep Fork, maybe Wanderer. Partly because of experiments in how far to stretch it, and I found these bots to be pretty mean in their tactics. I'm still fairly confident that I can win any base-10 game, given enough time, but not in an 'experimental' frame of mind ;-)

crush them! we are rooting for you from spain ;)

Thank you, I can use all the spititual help I can get :)

I too am rooting for Christian to defeat all the programs. If I'm not going to be the author the program that defeats Christian (That would be a major feather in anyone's cap.), then I can draw comfort from the fact that Christian can defeat all programs and I defeated* Christian.

\* [once]