Interesting. Is that because having more options makes it easier for you to force your opponent into a fatal capture sequence where he must capture all of your pieces?syzygy wrote:That worked in the first few days after suicide chess was added to FICS, but nowadays it will make the engine lose miserably.Evert wrote:Do you really need a different search for suicide chess? I never really looked at it because I don't find it particularly interesting, but I figured that making the piece values negative would do it?
In suicide chess mobility is all important, so the more pieces the better unless tactics tells you otherwise.
I had another look at it, and I think I understand how it works in principle, but I'm not entirely clear on how to implement it for a chess program. It would probably have been easier for me to understand 15 years ago before I wired my brain to think in terms of iterative deepening alpha/beta search. Let me see if I understand it correctly, in decreasing order of confidence on my part:What works extremely well in suicide chess to cover tactics is proof number search. But it's not trivial to combine this with alpha-beta.
- PN search answers a yes/no question about a node, like "is this node won?", where it will return "yes", "no" or "don't know/maybe". The latter only if you truncate the tree in some way (say, memory limits, or limits on the maximum search depth). In this sense it acts a bit like a bit-base.
- Instead of "won" (say) you can also backup the exact score (won-in-N) and line of play, so it can also tell you a line of play that leads to the score, in the same way as alpha-beta.
- The returned solution is not necessarily optimal in the sense of being the fastest win. In alpha-beta terms the returned score is a lower bound.
- If you ask "does this node lead to a forced mate?" you can backup the +mate/-mate score and get a variation out of it, however because of the previous point you cannot know if the other side doesn't have a faster mate that invalidates the one you just found.
I might play around with this for the purpose of building a mate-solver, which is fairly useful in Shogi and other drop games, but it may be interesting for fortress detection as well.
Impressive!Anyway, suicide chess is about to be solved as a win for white (only 1...b6 is not yet known to lose to 1.e3!).
Do you think there are any zugzwang positions among the starting positions for Suicide960 (ie, positions that are won for black)?