> It seems that you could work on all the tables simultaneously, in a > time-sliced sort of way, and bound your worst case to N times the > ideal case. If you do the same amount of work per time unit on each > pair of tables, then you'll finish up the easiest pair first, without > having to know which one it will be ahead of time. From there the > rest of the work becomes easier.
Over at IDSIA, they'e working on something like this for AI: the general idea is that if one is looking for a good "predictor", and is confident that the problem one has is better described by code than data, it should be possible to try interleaving all possible predictors, to find a "short" one with some acceptable overhead. -Dave
