Author: Bas Hamstra
Date: 09:17:47 06/01/00
Go up one level in this thread
On June 01, 2000 at 10:40:28, Brian Richardson wrote: >On June 01, 2000 at 09:17:59, Bas Hamstra wrote: > >>On May 31, 2000 at 19:24:37, Brian Richardson wrote: >> >>>On May 31, 2000 at 18:02:42, Bas Hamstra wrote: >>> >>>I did some quick tests with Tinker (of course, the search "ecosystem" of each >>>program will be very different). Normally, Tinker probes the hash table, then >>>does the hash move (if any), then killers, then generates moves, and so on. >> >>I hope you DO captures before killers? Doing that, killers should be a big win. >>But note that if you already do history, that the extra gain of killers is far >>less, because both try to achieve the same. > >No, I don't generate any moves (captures or non-caps) until after killers (or >the hash move) are tried. However, I don't use killers (or hash move) in >q-search, just full-width. Wait, that's bad! You *have* to do captures before killers. To quote Bob Hyatt: "captures are better killers than killers", which I think is a fact. *At least* the winning captures before the killers. >>>For the following, one history table for both sides is always used. Tested on >>>five positions (opening, two from WAC, one from Fine and one BT2630): >>>1) Turning hash moves off--usually 2x slower, up to 10x slower, one test 5% >>>faster. >>>2) Turning killers off (keeping 2 moves)--very mixed results: same, or up to 30% >>>faster or up to 10x slower (Fine16) >> >>This 10x slower is counter intuitive to me. If you don't use history sorting >>killers should be a big win. If you do, it is a small win overall. > >It was a big surprise to me too. The tests above were all with history. >> >>>3) Turning off both--better than no killers, but worse than no hash move. >>> >>>To my surprise, Tinker's WAC 300 score went UP with no killers. Again, this is with history (disabling history slows everything down A LOT). I may try some >>>self-play tests with and without killers. >>> >>>Brian Richardson >> >>Maybe you could also include: a) History and no killers > >This was test 2 above > >>and b) Killers and no history? > >No history is MUCH slower. > >>I suspect that if you already use 2 killers the gain of adding history >>is small. And since killers are cheaper I don't use history sorting... >> >> >>Regards, >>Bas. > >What I'm starting to see is that history is better than killers. The history >overhead is very small, since it is part of the move ordering score along with >MVV/LVA anyway. > >Some early self-play tests showed killers better at time 2 1 and no killers >better at 5 0. Will run some more tests tonight. > >Brian I think you should redo all tests doing captures first. It should give much different results. And then the question remains: which is most effective: 2 killers or history sorting, if you had to pick one? Also I think the overhead of killers is smaller than history. At least the way I do it. I don't use counters for killers. I just update KillerMove[Ply] whenever there is a move > alpha. Now KillerMove[Ply] is the first move to try after the (winning) captures, if it is legal. Bas.
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.