GameKnot: club forum - Using Engine Analysis Output Sensibly

« TAP TO LOG IN

Play online chess!

Using Engine Analysis Output Sensibly

« Back to club forum

From	Message
bigpeta 08-Aug-13, 06:41	Using Engine Analysis Output Sensibly Let me reiterate what I have said before about the use of engine analysis that you have not done for yourself. Then I will add a few more technical facts that previously were not necessary to include as they just complicate the issues. When an engine is used for analysis it scores moves in centipawns. ie one hundreth of a pawn. So in the scores you will see such numbers as 1.35. This means a whole pawn plus 35 centipawns. If the score is preceded by a minus sign then the advantage is in black's favour otherwise it is for white. This score is affected by many variables which can be set internally in the engine. The main ones are:- 1: Search depth - usually given as ply or half moves. There are three main numbers to bear in mind. 10ply will find most blunders 15ply will be good enough to beat most online players if used in a game. 20 ply will almost guarantee that the engine has found the best move. I have rarely seen an engine change its mind after 20ply. The danger with 10 or 15ply is the horizon effect. This where a move looks good/bad at 15 ply but searching deeper will show it to be the opposite. A good example is a major piece sacrifice which requires say 18 ply to reach mate. At 15 ply it may still look like a blunder. 2: Engine strength and attitude. An engine can be set to play at any elo strength you wish plus you can tell it to be aggressive or defensive. Both of these settings will affect how the engine assesses a position and that could result in a different score. It could also result in a different move being suggested as an alternative. 3: Trigger values. This is the extra technical information I mentioned at the start. In the GK engine output you will see some moves marked as a mistake/error/blunder. How it assesses these are through parameters set when the analysis is run. It is told to treat specified values as triggers for the three levels of comment. for example it might set the mistake level as a score of 0.2. This means that if your move alters the score by at least this amount then the analysis will indicate a mistake and give the 'best' alternative. Similarly other triggers will be set for error and blunder. Typically these would be around 1.0 for error and 3.0 for blunder. That is a pawn or a minor piece. Of course you have to remember that the accuracy of the scores being used depends upon the ply level and other parameters that have been set. Also the relative seriousness of the error depends on the rating level of the players. For lower rated players (say < 1600) then the loss of a minor piece may only be a mistake. Whereas for higher ratings (say >2200) then even the loss of a tempo could lose the game. To show the differences I have posted my own engine analysis chart using Rybka 4 running under Shredder 12. elo set to max, attitude to neutral and ply to 20 plus. Because I cant post pictures here I have included a link to it on my my flickr account. For those of you not familiar with this type of chart then each bar represents a half move (1 ply). Green above the line is advantage to white. Red below advantage to black. The move numbers are shown along teh bottom. In the top left corner is the scal in centipawns. The 1st pic is my game with todd up to move 20. www.flickr.com As you can just see there are drops in my advantage score which were not enough to trigger the GK engine to comment. By changing the analysis to take in only the 1st 12 moves then these drops become more obvious. www.flickr.com Finally I have included the whole game analysis between my self and drhpatron. This game ended in a draw. As you can see there are many drops in advantage on both sides and yet the GK engine did not flag up a single one. The reason is that these changes were not large enough to trigger the alert levels set in the GK engine. www.flickr.com I hope that the notes above will help you to make better use of engine analysis output. As a final remark I will say that the analysis should be the start point for your own analysis and not the end. It is a tool to help you and not a replacement for your own thinking.
bigpeta 15-Aug-13, 07:41	Interpreting the engine output A question from greenways116 indicated that i should carry this look at engine output a little further. He asked how to interpret the alternative moves suggested by the engine. Before you read further try and analyse the 'best' move for black in the following position. First you need to understand an engine parameter PV (preferred values), this tells the engine how many alternative lines to display. These are in ascending order. The GK analysis only gives one alternative and this is supposed to be the 'best' move. As can be seen from the following example (based on engine analysis of the position above) this is far from the truth. The best move is Qb5+ but it only beats Rxd2+ by 1 centipawn. It also only beats the 3rd and 4th choices by 3 and 4 centipawns respectively. To all intents these 1st 4 options are equal and black could play any of them without causing a problem for his position. The only real change comes when we look at PV 5 and this shows a jump to over a whole pawn. So obviously this is not a good move to play. 1. +/= (0.31): 29...Qb5+ 30.c4 Rxd2+ 31.Kxd2 Qb4+ 2. +/= (0.32): 29...Rxd2+ 30.Kxd2 Qb8 31.Kc2 g6 3. +/= (0.34): 29...Qxe4+ 30.fxe4 Rb8 31.Rb2 f6 4. +/= (0.35): 29...Qa6+ 30.c4 Re8 31.Rd5 h6 5. +/- (1.18): 29...Qc7 30.Rxd8+ Qxd8 31.Qxe5 f6 So having decided on the best available moves what next? The important move is the 1st one suggested. What follows after that is conjecture by the engine about possible continuations. First you have to see if you can understand why these moves are better than your own. The suggested continuations can help you see where the engine is going. If you cant see the reason for the suggestion then that is the time to ask for help from your tutor/mentor. Next you have to assess how much better the suggested moves are than the one you played. There are 2 ways to do this:- 1: play your move and see how the engine scores it and check the difference. 2: Much better is to increase the PV count until your move appears on the list. But higher PV and deeper ply count will soak up lots of time unless you have a fairly powerful computer. The deep versions of the engines will use as many cores and as much ram as you can provide. The alternative is to let it run over night. As an example my fast m/c with a quad core and 8Gb ram took 15 minutes to take the above PV 5 to 20 ply. One benefit of the 2nd method of increasing the PV count is that when your move appears it gives an indication of how good your analysis is. You should be looking to have your choice in the top 3 lines. (that means the 3 best and in this case ties making best 4 lines) As a rough guide if you are over 1900 you should be in the top 3 lines. if you are 1600 to 1900 then the score for your move should be no worse than 1 pawn below the best. below 1600 you should be looking to be no worse than 2/3 pawns worse than the best. These are only guidelines and not rules written in stone.

From

Message

bigpeta
08-Aug-13, 06:41

Using Engine Analysis Output Sensibly

Let me reiterate what I have said before about the use of engine analysis that you have not done for yourself. Then I will add a few more technical facts that previously were not necessary to include as they just complicate the issues.
When an engine is used for analysis it scores moves in centipawns. ie one hundreth of a pawn. So in the scores you will see such numbers as 1.35. This means a whole pawn plus 35 centipawns. If the score is preceded by a minus sign then the advantage is in black's favour otherwise it is for white.
This score is affected by many variables which can be set internally in the engine. The main ones are:-

1: Search depth - usually given as ply or half moves.
There are three main numbers to bear in mind.
10ply will find most blunders
15ply will be good enough to beat most online players if used in a game.
20 ply will almost guarantee that the engine has found the best move.
I have rarely seen an engine change its mind after 20ply.
The danger with 10 or 15ply is the horizon effect. This where a move looks good/bad at 15 ply but searching deeper will show it to be the opposite. A good example is a major piece sacrifice which requires say 18 ply to reach mate. At 15 ply it may still look like a blunder.

2: Engine strength and attitude.
An engine can be set to play at any elo strength you wish plus you can tell it to be aggressive or defensive. Both of these settings will affect how the engine assesses a position and that could result in a different score. It could also result in a different move being suggested as an alternative.

3: Trigger values.
This is the extra technical information I mentioned at the start.
In the GK engine output you will see some moves marked as a mistake/error/blunder. How it assesses these are through parameters set when the analysis is run. It is told to treat specified values as triggers for the three levels of comment.
for example it might set the mistake level as a score of 0.2. This means that if your move alters the score by at least this amount then the analysis will indicate a mistake and give the 'best' alternative.
Similarly other triggers will be set for error and blunder. Typically these would be around 1.0 for error and 3.0 for blunder. That is a pawn or a minor piece.
Of course you have to remember that the accuracy of the scores being used depends upon the ply level and other parameters that have been set.
Also the relative seriousness of the error depends on the rating level of the players. For lower rated players (say < 1600) then the loss of a minor piece may only be a mistake. Whereas for higher ratings (say >2200) then even the loss of a tempo could lose the game.

To show the differences I have posted my own engine analysis chart using Rybka 4 running under Shredder 12. elo set to max, attitude to neutral and ply to 20 plus.
Because I cant post pictures here I have included a link to it on my my flickr account. For those of you not familiar with this type of chart then each bar represents a half move (1 ply). Green above the line is advantage to white. Red below advantage to black. The move numbers are shown along teh bottom. In the top left corner is the scal in centipawns.

The 1st pic is my game with todd up to move 20.
www.flickr.com
As you can just see there are drops in my advantage score which were not enough to trigger the GK engine to comment.

By changing the analysis to take in only the 1st 12 moves then these drops become more obvious.
www.flickr.com

Finally I have included the whole game analysis between my self and drhpatron. This game ended in a draw. As you can see there are many drops in advantage on both sides and yet the GK engine did not flag up a single one. The reason is that these changes were not large enough to trigger the alert levels set in the GK engine.
www.flickr.com

I hope that the notes above will help you to make better use of engine analysis output. As a final remark I will say that the analysis should be the start point for your own analysis and not the end. It is a tool to help you and not a replacement for your own thinking.

bigpeta
15-Aug-13, 07:41

Interpreting the engine output

A question from greenways116 indicated that i should carry this look at engine output a little further.
He asked how to interpret the alternative moves suggested by the engine.

Before you read further try and analyse the 'best' move for black in the following position.

First you need to understand an engine parameter PV (preferred values), this tells the engine how many alternative lines to display. These are in ascending order.

The GK analysis only gives one alternative and this is supposed to be the 'best' move.
As can be seen from the following example (based on engine analysis of the position above) this is far from the truth.
The best move is Qb5+ but it only beats Rxd2+ by 1 centipawn. It also only beats the 3rd and 4th choices by 3 and 4 centipawns respectively. To all intents these 1st 4 options are equal and black could play any of them without causing a problem for his position.
The only real change comes when we look at PV 5 and this shows a jump to over a whole pawn. So obviously this is not a good move to play.

1. +/= (0.31): 29...Qb5+ 30.c4 Rxd2+ 31.Kxd2 Qb4+
2. +/= (0.32): 29...Rxd2+ 30.Kxd2 Qb8 31.Kc2 g6
3. +/= (0.34): 29...Qxe4+ 30.fxe4 Rb8 31.Rb2 f6
4. +/= (0.35): 29...Qa6+ 30.c4 Re8 31.Rd5 h6
5. +/- (1.18): 29...Qc7 30.Rxd8+ Qxd8 31.Qxe5 f6

So having decided on the best available moves what next?
The important move is the 1st one suggested. What follows after that is conjecture by the engine about possible continuations.
First you have to see if you can understand why these moves are better than your own. The suggested continuations can help you see where the engine is going. If you cant see the reason for the suggestion then that is the time to ask for help from your tutor/mentor.
Next you have to assess how much better the suggested moves are than the one you played.
There are 2 ways to do this:-
1: play your move and see how the engine scores it and check the difference.
2: Much better is to increase the PV count until your move appears on the list. But higher PV and deeper ply count will soak up lots of time unless you have a fairly powerful computer. The deep versions of the engines will use as many cores and as much ram as you can provide. The alternative is to let it run over night.
As an example my fast m/c with a quad core and 8Gb ram took 15 minutes to take the above PV 5 to 20 ply.

One benefit of the 2nd method of increasing the PV count is that when your move appears it gives an indication of how good your analysis is.
You should be looking to have your choice in the top 3 lines. (that means the 3 best and in this case ties making best 4 lines)

As a rough guide if you are over 1900 you should be in the top 3 lines.
if you are 1600 to 1900 then the score for your move should be no worse than 1 pawn below the best.
below 1600 you should be looking to be no worse than 2/3 pawns worse than the best.
These are only guidelines and not rules written in stone.

« Back to club forum

GameKnot: play chess online, chess clubs, Internet chess league, monthly chess tournaments, chess teams, online chess puzzles, free online chess games database and more.