Indeed, the number of required games increases fast. But the WinBoard games are not so much a problem, they can be run automatically overnight. So Zillions and ChessV are the participants that really count, in terms of effort. Perhaps it is an idea to first play Mat's improved version of Zillions against all the other engines. Then, if it loses all games, it might be pointless to play the weaker version of Zillions. ;-)

Btw, Gregory Strong is working on the follow-up of ChessV. It is called Quadrox, and it already plays normal Chess. It participated in my on-line blitz tourney yesterday. It is planned to be a WinBoard engine as well, and soon will be able to do 10x8 Chess.

Not all engines are deterministic, and Joker80 is mildly non-deterministic. (At very long time control it might become more deterministic, though.) It adds a pseudo-random term to its evaluation, determined from the starting time of the game and the hash signature of the position. This way of implementing randomization acts as a kind of poor-man's mobility evaluation, and thus should increase playing strength compared to the purely determiistic version.

Another method of introducing randomness is to add a small pseudo-random number to the score of every move in the root. When several moves have nearly equal score, this should randomly choose between them. This does not increase playing strength, but it stays effective at arbitrary large search depth.

Put this multi-PV mode (givin scores for several moves, rather than just the best) such as you show on ChessV can definitely be very useful for constructing a book. Perhaps I should implement it in Joker80 too. The idea of a book is that you are willing to play theoretically sub-optimal moves to provide diversity, so that the opponent cannot easily prepare, and thus will eventually play even more sub-optimal moves. If you are striving for a variabilty of 1000 lines, and you can choose between 10 moves on each ply, you only need a book of depth 3. (Well, perhaps 4 because many lines are transpositions of other lines.) But you would be choosing the 10th best move on some of these lines, and its score might be really a lot lower than that of the best. So in practice you only choose on average between 2-3 moves on each ply, and your book had to be 10 deep before it fans out to 1000 lines (assuming the opponent always plays his best move). But then all these lines can be very close in score to the best line. So there is a trade-off between depth of the book and how much score you have to sacrifice to get a given variability.

It would be interesting to develop (and implement) an algorithm that would create the minimal tree that would have a given number of end leaves with a score (as determined by an engine search of a given quality) within a preset margin from the best line. E.g. how much variation would it buy us if we are willing to sacrifice 30 centiPawn after 1, 2, 3, 4, 5 moves, against optimal play.