lichess.org
Donate

Question about SF Analysis

Is the generated move values same for everyone?

because i was analyzing with a friend, her values were quite different than mine

for example, in the image below you see that the move value for black changed from +1.3 to +01 to my advantage

http://i.imgur.com/yKSPTHv.png

yet it wasnt the same value for my friend, it was +1.7! thats a big difference and there was no way of telling which value was correct

is this issue related only to antichess?
In general, there are several factors that can affect a chess engine's evaluation from run to run.

A first major one is depth. If one of you is getting a depth 24 max search, and the other a max 18 depth, then the difference could just be from one seeing more deeply.

A second one is parallelism. Parallel implementations in chess engines are very sensitive to timing, so for practical purposes running analysis with >1 core makes the results of search non-deterministic.

In a very complicated, tactical position with big swings in evaluation as the search goes deeper, the above factors can easily result in big differences for two different runs.

For what it's worth, I get 0.0 for that position myself.

@a_pleasant_illusion Thank you for your reply. I think i understand what you mean. My friend had the same depth "24"

how do you know which is more accurate evaluation?
@Dragon-Lord Since both analyses have the same depth, neither is more accurate than the other. Just as a comparison, if you ask two 2500 rated players to evaluate the position, how do you know which evaluation is more accurate? You would have to analyze the given lines to get an idea of which one might be more accurate and this applies to both the humans' and the engine's results.

As @a_pleasant_illusion explained, the non-deterministic result is caused by multithreading.
Indeed, evaluations are nondeterministic and the entire concept of measuring CP loss is more tenuous than most people assume it to be!

This topic has been archived and can no longer be replied to.