There is no reason to believe the lines will cross because they seem to be moving away from each other starting at the origin.
Also, the lines are not straight, there appears to be a gentle slope to both of them. A decade or two ago we talked of these lines being straight or "linear" because we could only see a portion of what we can see now. It was not feasible to run thousands of games at high depth levels. I'm doing more testing in one day than we could do in 2 or 3 years. It's possible that there would be less curve if I plotted by time or if I used a full width version of both programs. I can only guess about this. I believe it's the case that a good selective search improves faster with time, but slower with depth. The metric that's most important for practical engineers is time because you want a program to play as well as possible in the least amount of time. I would have preferred to do this study with time, perhaps setting up levels that used 2X more time than the previous. However it's difficult to do this accurately on a computer that is being heavily used for other things too. It would have been possible to do this based on total nodes searched but that would have required some re-engineering of the autotester (to make it abort searching when the specified number of nodes have been exceeded and to ignore moves that might have been selected after this.) My own chess programs always had a way to specify levels based on total nodes searches because it was very useful for consistent testing. If you look at the chart, you see that in a sense the versions do "cross" each other, at least at the moment and based on the current data which is still subject to a lot of statistical noise. Consider what you see now if you assume the numbers are accurate: For the first 3 depths, the program that is doing the deeper search is the strongest regardless of evaluation function. For instance Weak-3 is stronger than Strong-2, just as you would expect. However, after that it appears that the weak evaluation function needs 2 extra ply to be stronger. At the high end of the range, it appears to require an addition 3 plies of depth to be stronger because a 9 ply search with the strong evaluation function is beating the weak 11 ply search and so it would require at least a 12 ply search with the weak evaluation function to beat a measly 9 ply search. If those numbers hold up, the implication is that the deeper you search, the more important your evaluation function becomes. I'm sure it's possible to put other spins on this and I'm sure people will, but this seems like the most reasonable working premise. I think the numbers will hold up even though there is a lot of noise in the data. Although each data point is noisy, the general trend is not ambiguous. - Don Ivan Dubois wrote: > I have a question : > If the lines in the graph are straight lines and they dont have the > same increase rate, then isnt there a point where they should cross ? > Do they all cross at the same point ? > I guess this point (if it exists) would indicate some kind of starting > point : It would correspond to the weakest possible strenght. > > Any thoughts ? > > _______________________________________________ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/ > _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
