PDA

View Full Version : Why does Fritz change it's analysis within a move if you let it play itself?



Paul Cavezza
10-11-2009, 03:32 AM
Oddly phrased question but i'm tired. What i'm asking is why Fritz will give a rating like +=.89 after 4 hours thought, you get it to play the +.89 move and then switch sides and think for 4 hours and within a move or two the reading is changed. What is the point of these numbers if they can change so quickly. To what extent do top players trust computer analysis?

michael.mcguirk
10-11-2009, 05:46 PM
I would think the obvious answer is that the depth of the calculation that it follows. While the tactical responses would be realised by Fritz to their end, any positional values would only be realised to a certain point, after which the move is played, it goes half-ply deeper. I could be wrong, but that was my understanding of it.

Kevin Bonham
10-11-2009, 09:57 PM
I find that leaving Fritz to run on infinite analysis on one move for a very long time is not a very productive way to get an accurate evaluation for a position. Sometimes it will still miss something (a long-term sacrifice especially) that it might quickly pick up on as a good move once a few moves from the line are entered in. Also, some lines have a tendency to blow out in particular directions - this especially applies to positions with compensation for material deficit. Thus a bit of "prodding" leads to a better outcome. I've noticed that Garvin (who is significantly more successful at freestyle than I am) has made similar comments about the need to encourage the engine (in his case Rybka) to consider particular ideas.

Garvinator
10-11-2009, 11:55 PM
Thus a bit of "prodding" leads to a better outcome. I've noticed that Garvin (who is significantly more successful at freestyle than I am) has made similar comments about the need to encourage the engine (in his case Rybka) to consider particular ideas.
This is correct. If players could just let their engines run for a day or so and have perfect play, then anyone could be good at freestyle and every game would be decided either by opening prep alone or would be a draw. Of course this is not the case.

The reason the engine changes its evaluation is two fold:

1) Depth of analysis as in how many ply ahead the engine is evaluating. The engine will change its evals if it discovers something (good or bad) in the next set of ply.
2) Hash tables. Each engine uses hash tables to store previous information to save on processing time and to ensure the engine does not have to keep repeating the same sections of analysis over and over again. The key to using hash tables correctly even if you let the engine run on infinite analysis with no prompting is to make sure your hash tables are large enough that your engine can store a lot of previous analysis, but not too large that it slows up the computer and freezes the memory.

Engines also suffer from the 'horizon' effect, where if a good move is at the end of a line of analysis, but is outside of the range of the engines analysis ability, then it will not 'see' the good move and will continue down this line of analysis until it is too late.

This is where human intuition and some prodding is required. The human senses that there is an attack of some kind and starts making moves, then going back and making a few more moves. The evals change as this is done because the human is 'helping' the engine to fill up the hash tables and to extend the horizon effect of the engine. It also speeds up the early analysis of the engine.

ER
11-11-2009, 07:05 AM
In other words "I have a bigger/better one. I win"! That's why forget about Corr Chess and concentrate in real life OTB chess! ok I will ask Fludey in the interview before I commit myself into http://www.pic4ever.com/images/2mo5pow.gif and http://smileys.on-my-web.com/repository/Computer/computer-18.gif

Desmond
13-11-2009, 09:23 PM
It's probably a female.

Davidflude
13-11-2009, 11:37 PM
Saw the claim on a site that I regard as reliable that you can tell in high level correspondence play as to which players use computers in endgames because their results tend to be worse.