Author: Ratko V Tomic
Date: 07:00:22 11/19/99
Go up one level in this thread
> * Tiger's book learning is almost nonexistent. Tiger just avoid > repeating lost games. How far back in the lost game does it start the avoiding? With Fritz 5.32, after it lost few games in an opening (for reasons well beyond and unrelated to the opening), it stopped playing a perfectly good entire opening completely. I had to reinstall the book from the CD (since it wouldn't reset the weights properly on the copy it had on the hard disk). And after that would give only a short lived variety in openings (no matter what its book menu options said, it seems to learn its own way, anyway) I had to reinstall it and make the book read-only file to get rid of its excess in cleverness. Similar problem occured with its forcing winning lines, where it started playing the same opening line repeatedly, even though its wins again had nothing to do with the opening but came much later (mostly for resons of tactical oversights). So Fritz (or Hiarcs) learning is basically rigged to squeeze an extra point or two in mindless comp-comp marathons against the less aggressive learners, at the great inconvenience of its end user (but the extra few points from the mechanical comp-comp matches give it a greater attractivenes to the potential new customer, i.e. CB's customer philosophy appears indistinguishable from that of a used car salesman). For program-human play, the avoiding should not banish entire good opening just because program lost later for unrelated reasons to a stronger or more inspired opponent. The only thing (useful to an end user) to avoid shold be moves producing explicitly the positions where the programs score dropped significantly (as evaluated during the losing game, provided time setting was not lower than the current one). Optimally, such positions to avoid could be part of the hash tables so they can be detected during the search as early as possible. A small subset of them need to be loaded at any one time, for different openings or some other (more reliable) classification e.g. positions with more than Nw or more than Nb pieces need not be loaded, or can be unloaded if already loaded, if white has lower than Nw pieces, or black has lower than Nb pieces (or they can be even subdivided by piece type to minimize the memory requirements). In any case, such long-term position memory would provide a set of nodes evaluated typically better than what the program would do in a new search (even with higher time/move or depth settings, since these would be found at some depth in a new search tree). The main advantage of such learning method for the user isn't so much that the program will play stronger as it remembers more pre-evaluated nodes (even though that would be one side-effect), but that the program would not waste its user's time "thinking" for minutes about the same position it "thought" about for the same time, perhaps yesterday, producing the same move (whatever the game outcome might have been). That's one quite irritating thing about the programs (for people who have something else to do besides playing against the program), and I still haven't seen a commercial program which will do it, but without ending up playing the same opening line over and over (i.e. the CB's 'used car salesman' learning system). > Tiger is unable to repeat a won or a draw game, > except by pure luck. I admit this is maybe a mistake, I might change > my mind in a future version. If you add that kind of learning, it would be nice if it were separate option from the avoidance of lost positions, since it automatically leads to narrowing of the opening book (or rather, the non-uniform weights result in the lower entropy/information content of the opening repertoire; the entropy is maximized by uniform weights on all book lines). There is no reason (other than either the salesman mentality taking over the design or the programming laziness/incompetence taking shortcuts) to bundle inseparably the options best for the mechanical comp-comp marathons, with those optimal (not just in strength but also in saving user's time and being pleasant, having variety) for play against the human end user. An option which explicitly sets the Strongest play should be split to Strongest against humans vs Strongest against computers. While Rebel 10 does have something of this sort (in an unclear and somwehat overlapping indirect ways, via anti-GM vs Tiger vs Strongest Settings vs 10b vs 10C options), that still refers to the playing/searching style, not to the book (or rather, position) learning as described above.
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.