Author: Bas Hamstra
Date: 21:56:07 01/12/01
Go up one level in this thread
On January 12, 2001 at 22:59:15, Dann Corbit wrote: >On January 12, 2001 at 22:46:33, Bas Hamstra wrote: > >>Within a few days I will put a new version of Tao 5.5 on Frank's web site. There >>have been several small improvements, resulting in an ELO improvement of exactly >>50.00 points. The main difference is that it uses evaluation learning based on >>Temporal Difference. With all parameters set to zero it quickly learns how to >>improve the most important eval terms, and play better chess. In my opinion >>however, it is not able to set all terms to optimal values within a reasonable >>amount of time. Judge it yourself, you can let it learn from zero, or you can >>turn learning of and tweak the weights manually. So far, it seems best to start >>with a good guess and hand-tweak the few terms it does not seem to learn (or >>learns too slow). Having learning turned on has no performance penalty. >> >>The main differences with the previous version are: >> >>- Temporal Difference learning (can be scaled from 1000 to zero = off) >>- More accurate repetion detection >>- Better tactics (290 WAC within 5 sec) > >What machine are you using? >The only programs that do better than that on my machine at that time control >are Crafty 18 and ExChess 4. Both of those get 292. My machine is a 1GHz >Athlon (980 MHz, actually, I think). I use a Celeron 466 which is anything but ideal for this engine, as I noticed on the dutch open, where I used a PIII-450 laptop. Together with a VC compile (my work compiler is Borland BCC) it did 2x the NPS and went from 125K to 250K. Amazing, because I didn't expect much good from that laptop. There was hardly anyone that outsearched me, to my surprise. >>- Eval changes to let it play more interestingly and actively >>- Eval parameters user settable >>- More accurate PV's (less half PV's that end in the hash table) >>- Improved search efficiency >>- Some changes and additions to the evaluation >> >>I have not hand tuned the eval weights for a long time, like in the previous >>version. Therefore the eval is sometimes over-optimistic and maybe not totally >>balanced. But still, IMO, it plays very interesting attacking active REAL chess. >>I have not had time to test it at very long time controls. > >If you can send me a copy right away, I will run it over the weekend on a >of fast machines. Otherwise, I can wait until it get posted to Frank's page. That's great, I will try to send it tomorrow. Ciao! 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.