Marc Benford wrote on 07/03/14 at 20:23:26:

They are not in the same units.

Quote:

What do you mean by "accurate"?

You should redo the poll with lots more significant figures, I don't think that it's precise enough as-is.

Quote:
No need to go up to depth infinity. Depth 11,741 is good enough (it's the maximum amount of plies theoretically possible if I am not mistaken).


Quote:
I think the definition of what I meant to ask was definition *D*.


Quote:
Yes you're probably right.



I think it would be interesting if some people who have lots of engines let their engines analyze the starting position and post what evaluation each of them give.

Then we could make the average of all these evaluations, and we would have a good approximation of the true evaluation of the starting position.

I just made my engines evaluate the starting position.

Houdini 3
Depth = 26/65
Time = 70 minutes
- 1.Nf3 : +0.20
- 1.d4 : +0.20
- 1.e4 : +0.15
- 1.e3 : +0.13
- 1.c4 : +0.11
- 1.Nc3 : +0.10
All 14 other moves are +0.03 or less.

Fritz 12: 
Depth = 23/54
Time = 140 minutes
- 1.e4 : +0.53
- 1.d4 : +0.34
- 1.Nf3 : +0.29
- 1.Nc3 : +0.12
- 1.e3 : +0.04
- 1.c4 : -0.01
All 14 other moves are -0.03 or less.


Edit: Okay I just let Fritz evaluate the starting position for more than 10 hours. Here's the result:

Fritz 12: 
Depth = 25/65
Time = 10 hours and 30 minutes
- 1.d4 : +0.38
- 1.e4 : +0.32
- 1.Nf3 : +0.30
- 1.Nc3 : +0.14
- 1.e3 : +0.08
- 1.g3 : +0.05
- 1.c4 : -0.07
- 1.a3: -0.13
- 1.c3 : -0.16
- 1.h3 : -0.17
- 1.d3 : -0.20
- 1.f4 : -0.23
- 1.b3 : -0.25
- 1.a4 : -0.42
- 1.h4 : -0.47
- 1.b4 : -0.47
- 1.Na3 : -0.52
- 1.Nh3 : -0.56
- 1.g4 : -0.77
- 1.f3 : -0.90

JohnG wrote on 06/29/14 at 17:50:23:

. .
Pale Horse, Pale Rider wrote on 06/29/14 at 20:16:37:

Exactly.
Marc's question is not stupid, but it is poorly phrased.
This is not a question for computers, it is merely a dictionary question that is stated in an obfuscated way.
Correctly rephrased, Marc's question would be more like the following:

What should a chess playing software program strive to evaluate?


*A* Should it measure, separately for each player, the breadth of the variation tree that does not lead to a loss, and report the ratio as the evaluation, on the idea that the less broad your safe move options are the more likely you will make a move that strays outside those options?

*B* Should it statistically measure the actual outcomes, summed from large databases, for "similar" positions and use the historic percentages of wins and losses as the eval?

*C* Should it estimate the percentage chance that the player with the advantage will win? (I think this is what Rybka 4 does.)

*D* Should its evaluation foundation be the materialistic centipawns, where various positional factors are assigned a value as fractional pawns? (This is what Fritz 11 does; although Fritz 14 might have switched to a different eval?).

*E* Should it strive to fully calculate all 10^33 moves among all variations to prove whether White or Black has a forced win, and if neither has then evaluate as exactly 0.0?


There is no practical doubt that - Black can hold a draw if Black plays perfectly. Yet databases prove White wins more than Black does.
So Marc...

When you settle on a definition of exactly what you mean to ask, the answer will become easy for all to see.
But then, others might debate whether alternative definitions are more useful.
. .

This is a question for computers, not humans. 

I'll wait for some responses, and then consider moving it to the Chess and Computers section.

Stupid.