Author: James Swafford
Date: 09:31:55 08/07/02
Go up one level in this thread
On August 07, 2002 at 11:49:13, Jim Monaghan wrote: That's an interesting test... and it makes sense to me, too, given that search depth increases exponentially. So, as time increases, your "delta depth" decreases. If you put time on the X axis and rating on the Y axis, it (sorta) resembles the log function. There seems to be an upper bound on the rating, so that the graph would flatten out and hug that upper bound as time increases to infinity. And it's not defined for negative time. -- James >Hi, > >I've been playing around with some position testing and engine scoring on epd >tests. I've come up with the following table which expresses a person's (or >engine's) rating based on time given for thought. Here it is: > >Time (sec) Rating >1 2181 >2 2294 >3 2334 >4 2377 >5 2394 >6 2417 >7 2434 >8 2451 >9 2461 >10 2464 >11 2477 >12 2484 >13 2491 >14 2501 >15 2504 >20 2514 >25 2531 >30 2534 >40 2547 >50 2564 >60 2571 >90 2581 >120 2604 > >A number of things can perhaps be gleaned from the above. > >1) After about 15 seconds there is an appreciable slowing of ascent of the >rating number. A kind of diminishing returns sets in around here. >2) For engine matches, this perhaps means that reasonable testing results could >be achieved with a time control of around 15 seconds per move (on good hardware) >if time is a factor. Time efficiency is usually a factor. Assuming a typical >computer game lasts about 60 moves between competent programs, that's about Game >in 15 minutes. >2) There is still a significant difference between 2504 (at 15 seconds per move) >and 2604 (at 2 minutes per move). >3) One could surmise that a 2604 player, playing eight 2504 opponents >simultaneously to avoid the pondering factor would score 50% (baring tiredness). >The assumption is that the 2604 would be spending 15 seconds per move on each >game, while the opponents would be thinking for 2 minutes per move on average. > >I apoligize if this kind of thing been presented before. It seems to make sense. >A question for the mathematicians here. What is the formula that can be used to >express this time/rating relationship? It makes a pretty nice ascending graph. >Remember also, that it's the differences between the ratings that are important >rather than attaching any significance to the absolute numbers themselves. > >I derived this data from running Shredder Paderborn through the IQ test. The IQ >formula is IQ elo = 1564 + (% score x 12). So in the above table for example, >Shredder scored 75.0% at 10 seconds for an IQ elo of 2464. It's IQ elo at 2 >minutes was 2604 based on a score of 86.67%. PC used is a Cel 1.3 Ghz. > >Cheers, >Jim
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.