Evaluating my evaluation

Discussion of chess software programming and technical issues.

Moderators: hgm, Harvey Williamson, bob

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Post Reply
sandermvdb
Posts: 136
Joined: Sat Jan 28, 2017 12:29 pm
Location: The Netherlands

Evaluating my evaluation

Post by sandermvdb » Tue Jun 12, 2018 6:40 pm

Below is a list of all my evaluation terms including the error when this specific term is disabled. The error is calculated using the Texel's tuning method using a quiet set (no winning captures possible). The list shows in other words how important is a certain evaluation term?

Code: Select all

0,05045	 original	
0,05046	 Shield 3            	0,00000
0,05046	 Shield 0            	0,00001
0,05048	 KS rank             	0,00003
0,05049	 Bishop outpost      	0,00004
0,05049	 Passed candidate    	0,00004
0,05049	 Mobility knight eg  	0,00004
0,05049	 KS unsafe check     	0,00004
0,05050	 Mobility queen      	0,00005
0,05051	 Knight outpost      	0,00006
0,05051	 No minor defenders  	0,00006
0,05051	 PSQT queen eg       	0,00006
0,05052	 Hanging pieces      	0,00006
0,05053	 Mobility queen eg   	0,00007
0,05053	 Bishop pawn         	0,00008
0,05053	 PSQT bishop eg      	0,00008
0,05054	 KS queen tropism    	0,00009
0,05054	 Discovered          	0,00009
0,05055	 Pinned              	0,00010
0,05055	 Mobility king       	0,00010
0,05056	 Rook trapped        	0,00011
0,05059	 Hanging pieces 2    	0,00014
0,05059	 Pawn neighbour      	0,00014
0,05060	 PSQT knight eg      	0,00014
0,05060	 PSQT rook           	0,00015
0,05061	 Mobility rook eg    	0,00015
0,05061	 Mobility knight     	0,00016
0,05062	 Mobility king eg    	0,00017
0,05062	 PSQT rook eg        	0,00017
0,05063	 Pawn supported      	0,00018
0,05064	 Imbalance           	0,00019
0,05066	 Mobility rook       	0,00021
0,05068	 Night pawn          	0,00023
0,05069	 Shield 1            	0,00024
0,05071	 KS other            	0,00026
0,05072	 Mobility bishop eg  	0,00027
0,05073	 PSQT king           	0,00028
0,05074	 Shield 2            	0,00029
0,05075	 KS check            	0,00030
0,05081	 Pawn connected      	0,00036
0,05082	 PSQT bishop         	0,00037
0,05083	 Passed king         	0,00038
0,05085	 PSQT pawn eg        	0,00040
0,05085	 Mobility bishop     	0,00040
0,05086	 KS check queen      	0,00041
0,05087	 PSQT queen          	0,00042
0,05088	 PSQT knight         	0,00043
0,05106	 PSQT pawn           	0,00061
0,05130	 Individual scores   	0,00085
0,05130	 PSQT king eg        	0,00085
0,05138	 Threat scores       	0,00093
0,05149	 Passed multiplier   	0,00104
0,05220	 KS no friends       	0,00175
0,05228	 KS pattern          	0,00183
0,05231	 KS attacks          	0,00186
0,05292	 KS score            	0,00246
0,05749	 Passed score        	0,00704
0,13044	 Material            	0,07999
Some notes
- 'KS scores' and 'Passed scores' shows the totals of other eval terms (KS pattern, KS attacks, Passed multiplier, etc...)
- 'Individual scores' consists of several terms, like rook on open file, double pawn, isolated pawn, ...
- Material is by far the most important (as expected), 2nd is passed pawn, 3rd is king safety
- my pawn shield adds very less to the overall evaluation, especially file 0 (H) and 3 (E). I can see that sometimes strong engines take advantage of this so this term needs improvement

MOBMAT
Posts: 63
Joined: Sat Feb 04, 2017 10:57 pm
Location: USA

Re: Evaluating my evaluation

Post by MOBMAT » Sun Jun 17, 2018 3:43 am

do you have a write up as to what all the terms mean?

they sometimes mean different things to different people.

or is that giving away the farm? LOL
Vince S
Author of MOBMAT

"Reductions, extensions, and pruning, oh my!"

Ratosh
Posts: 31
Joined: Mon Apr 16, 2018 4:56 pm

Re: Evaluating my evaluation

Post by Ratosh » Sun Jun 17, 2018 4:05 am

Hope you don't mind answering a few of my questions:

- The used input for those values is the epd file from Zurichess tuner with 725k lines?
- Do you clean up the input in any ways?
- Have you tried any other input file?
- How did you tuned PSQ tables?

Thanks

sandermvdb
Posts: 136
Joined: Sat Jan 28, 2017 12:29 pm
Location: The Netherlands

Re: Evaluating my evaluation

Post by sandermvdb » Sun Jun 17, 2018 10:29 am

MOBMAT wrote:
Sun Jun 17, 2018 3:43 am
do you have a write up as to what all the terms mean?

they sometimes mean different things to different people.

or is that giving away the farm? LOL
I have a write up, my sourcecode :)

sandermvdb
Posts: 136
Joined: Sat Jan 28, 2017 12:29 pm
Location: The Netherlands

Re: Evaluating my evaluation

Post by sandermvdb » Sun Jun 17, 2018 10:34 am

Ratosh wrote:
Sun Jun 17, 2018 4:05 am
Hope you don't mind answering a few of my questions:

- The used input for those values is the epd file from Zurichess tuner with 725k lines?
yes
Ratosh wrote:
Sun Jun 17, 2018 4:05 am
- Do you clean up the input in any ways?
I skip the in-check positions because these positions are (almost) never evaluated in an actual game.
Ratosh wrote:
Sun Jun 17, 2018 4:05 am
- Have you tried any other input file?
no
Ratosh wrote:
Sun Jun 17, 2018 4:05 am
- How did you tuned PSQ tables?
How do you mean? These were tuned the same way as the other evaluation terms.

Post Reply