>> Anything can exist in a random game :-) > Occur yes, repeat forever requires very special situations.
That makes long repeats unlikely in any individual run, and infinite runs infinitely unlikely, unless we run into one of those very special situations. But how special do these actually need to be? One popular way of scaling up Monte-Carlo based engines seems to add more processing resources to explore more of the search space. Now, if you have a loop that requires n specific steps with at most m loop-breakers at each step, it isn't one of those very special cases that would force a bot into the loop. But if your engine explores more than (m+1)^n alternative runs, one of its simulations is going to run into the loop, even if that isn't likely for each individual run, right? And if the class of dangerous situations isn't quite as narrow as one might think at first, then random simulations overall are more likely to run into one of them, even if we don't use such a "beast" to start the simulation. But even short of infinite loops, very few of those finite repetitions are likely to occur in real games - some of them have and do, but most are a sign of unrealistic simulations (unsupported claim;-), which are likely to mess with the quality of evaluation results. Since finite bounds can take care of infinite loops, such degradations in quality are more worrying, at least to me. Especially if their origins are hidden somewhere inside statistics gathered from a huge invisible (unrecorded for efficiency reasons) set of random simulation runs (tuned via experiments). Perhaps one could record a histogram of simulation lengths at least, and stop worrying unless there happen to be lots of entries in the >361 area. Otherwise, results might be influenced by whether the artifical bound on simulation length happens to fall into a high or a low of overlength runs. Claus _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
