Chess2u
Would you like to react to this message? Create an account in a few clicks or log in to continue.

Chess2uLog in

pollis stockfish's eval changing by itself?

yes, fish's eval is automatic like the automatic cars
1
17%
no, it is fixed and non-changing like normal engines
5
83%
Total Votes:
6

descriptiona deep extrapolation about stockfish's top-notch evaluation ( Emptya deep extrapolation about stockfish's top-notch evaluation (

more_horiz


this is a technical meditation on stockfish's performance
after watching countless number of matches of fish

I have read so many posts which said the same thing about the evaluation of fish
that is something like the following statement

fish is very aggressive!

when other top engines show
the white is winning only by +0.3 point

fish's eval says something like
the white is winning by + 0.8point
which is over exaggerated and expanded quite a bit

that has something to do with contempt score and contempt value and being aggressive or defensive


***


but I didn't read a post which says more in-depth analysis of fish's eval (maybe ppl said about this many times before
maybe I just didn't find those post)

I think fish's evaluation code is quite different from other engine's code

what I mean by that to be exact is
fish's eval can be non-fixed and variable

I think other top engine's eval is kind of fixed and fixated
(I'm no professional on computers
I just read and studied some com books only for hobby
if my opinion in this post is wrong
you should correct me in every way possible
I'm just a normal person who read some com books only for hobby)


***


that's why we don't have to change the contempt value of fish engine manually cuz it is all done automatically!!!

like the automatic gear


like 20 years ago, when we had cars
those cars had non-automatic gear

but nowadays, I think most cars come with automatic gear
which is easier to handle


***


let's say
fish is having a match with another engine named desk (with elo of the same or similar strength)

when fish is winning
the contempt score is high
(higher than another engine's)
(aggressive mode enabled automatically)


that's why the eval shows
fish is winning by much bigger scores


if fish is losing a match with the engine named desk

fish's eval shows
fish is losing by -1.5 point
when desk's eval shows
fish is only losing by -1.0 point

I think cuz fish's eval is automatically changed into the state of "defensive mode"
that means low contempt score


and it is finely tuned I think
so we never have to do manual setting of contempt score for fish

we don't have to set the contempt value of fish manually
cuz it is like the automatic gear!
it is done automatically
and that is my opinion
after making a observation of countless number of matches of fish

so when fish is winning the round
it is super aggressive to the point where it is super attacking
to make the round a win for fish

and when fish is losing a round
it is defensive in order for fish to draw
so when fish is losing
the objective of the match is not winning
but trying to make a tie
just trying to draw thumb up 

descriptiona deep extrapolation about stockfish's top-notch evaluation ( EmptyRe: a deep extrapolation about stockfish's top-notch evaluation (

more_horiz
evaluation of stockfish is wrong, I changed to houdini 3 that is more accurated in the evaluation. With stockfish you can get an evaluation of + 1.50 or more and suddenly the engine drawed or lost!!!!

descriptiona deep extrapolation about stockfish's top-notch evaluation ( EmptyRe: a deep extrapolation about stockfish's top-notch evaluation (

more_horiz
@jeandis wrote:
evaluation of stockfish is wrong, I changed to houdini 3 that is more accurated in the evaluation. With stockfish you can get an evaluation of + 1.50 or more and suddenly the engine drawed or lost!!!!



(also continued from the 1st post up there)



yes, that sounds right I guess
I'm not too sure since I'm no expert in com, but

I'm of the opinion that
houdini's eval is perhaps the best and most sophisticated

yes, that (fish can show fish is winning by +1.50 score then
all of a sudden, fish can get a draw or a loss) is correct!!!

it happened occasionally but not so frequently
I mean such a senario (leading by +1.50 then losing or tying) happened occasionally with older fish engines
(like a few months old)

but with the latest engines like fish Oct 9th version
I watched many matches of Oct 9th
but it lost or drew after leading the match by +1.50 or +2.00 point a lot less
it does the aformentioned mistake a lot less

the latest fish is quite different from the fish engines a few months ago


***


the easy explanation on why the situation happened was
cuz fish is playing super duper hyper aggressive when it is winning

(such as by +1.50, +2.00 score)

when an engine is playing too aggressively (to get a win)
(or to make the round a win for itself)
the holistic power of an engine decreases quite a bit

and that aggressive engine is more prone to making errors
and erroneous eval and future movements

it is more susceptible to mistakes and
it has a tendency to have weaker defence along the way
and it has a predisposition to make the future movement
that is reckless


***


and there's a technical explanation on why the above mentioned situation happened occasionally or sometimes more than occasionally with older versions

cuz the engines are made by using alpha-beta algorithm

the eval of an engine must follow the notion
the eval must be in line with such methodology of alpha-beta algorithm

and this has something to do with selectiveness


the topic of selectiveness and being selective is of upmost importance ever since the inception of chess programming
due to the fact and owing to the fact that
ppl don't have infinite amount of computing power

the CPU's calculating power is limited and finite

so if in one given position
if there are 50 possible movements one engine can choose from

one engine chooses perhaps 10 best moves
(cuz to calculate all the future deep moves of 50 movements
is not possible, it will take so much time)

so the engine only choose 10 best moves

then make ten six
make ten possible choices six

it's the process of narrowing it down

then the engine makes six two

so at this moment
the engine only has to think about two possible movements
instead of 50 or 20 different movements

and I think fish is very selective
and more selective than other engines

fish narrows the possible future moves down a lot
a lot more than other engines do
cuz fish's eval is very good
only an engine with superb eval can do it while maintaining good elo

when the selectiveness is very high
the engine can search for deep moves and deeper moves
during the same amount of given time


so fish can search for 20 plies in advance in 1 sec
when other engines can only search for 17 plies in 1 sec

so fish engine is I think very selective
older fish versions (a few months ago) made such aforementioned mistakes occasionally but
the latest engine (stockfish Oct 9th) doesn't make that kind of mistake
often from my observation thumb up 





***


update:


so alpha-beta algoritha and selectiveness is a dual-edged sword

it is the headache and the best thing for coding

***

and to increase attacking power

an engine can have high selectiveness

but it is a trade off

to have high selectiveness and increased attacking power

it must lose some other things
like defence and such
so the defence and other things will decrease

the engine must find the fine line between the two points
and keep the balance and reach the equilibrium of eval

descriptiona deep extrapolation about stockfish's top-notch evaluation ( EmptyRe: a deep extrapolation about stockfish's top-notch evaluation (

more_horiz

a deep extrapolation about stockfish's top-notch evaluation ( Adfqer10


a deep extrapolation about stockfish's top-notch evaluation ( Qerqrq10


descriptiona deep extrapolation about stockfish's top-notch evaluation ( EmptyRe: a deep extrapolation about stockfish's top-notch evaluation (

more_horiz
I analysed stockfish's eval
and usually SF can search deeper than houdini

but around 43rd moves
when the king is attacked by the queen


SF's search depth was dramatically lowered

I think it is important!
I am of the opinion that
when SF's king is under dangerous attacks
when SF's king's safety is in jepordy
some other new functions of codes kick in

and they are in effect
to protect the king
and keep the king safer!
that's why the CPU must run more instructions
thus the search depth goes down
but I think the searched move has better quality!


1. e4 c5 2. Nf3 d6 3. d4 cxd4 4. Nxd4 Nc6 5. Nc3 Bd7 6. h3 Nf6 7. Be3 g6 8. Qd2 Bg7 9. O-O-O Rc8 10. Kb1 O-O 11. g4 a6 12. f4 b5

13. Nb3 Be6 14. Bg2 Na5 15. Bd4 Nc4 16. Qe2 b4 17. Nd5 {+0.34/15} Bxd5 {+0.13/13} 18. exd5 {+0.42/14} a5 {+0.17/15} 19. Rhe1

{+0.22/17} Re8 {+0.17/15} 20. Bf3 {+0.24/17} a4 {+0.25/15} 21. g5 {+0.20/18} Nxb2 {+0.15/15} 22. Bxb2 {+0.48/19} axb3 {+0.07/15}

23. cxb3 {+0.54/19} Nd7 {+0.16/16} 24. Bxg7 {+0.00/19} Kxg7 {+0.15/15} 25. Bg4 {+0.28/19} Rc5 {+0.13/14} 26. Qd2 {+0.14/16}

Nb6 {+0.13/14} 27. Bf3 {+0.12/16} Qc7 {+0.11/14} 28. Qxb4 {+0.24/16} Rc8 {+0.19/15} 29. Qd2 {+0.26/17} Qb7 {+0.19/16} 30. Qb4

{+0.00/17} Kg8 {+0.24/15} 31. Rc1 {+0.40/17} Qa6 {+0.24/16} 32. Qd4 {+0.60/16} Qb5 {+0.16/16} 33. Qe3 {+0.64/15} Qb4

{+0.24/16} 34. Rxc5 {+0.48/16} dxc5 {+0.21/14} 35. Rc1 {+0.66/16} c4 {+0.17/17} 36. Qd4 {+0.54/17} Qa5 {+0.07/17} 37. bxc4

{+0.00/15} Qb4+ {+0.06/17} 38. Ka1 {+0.00/20} Rxc4 {+0.17/18} 39. Rxc4 {+0.00/22} Nxc4 {+0.13/20} 40. Bd1 {+0.00/24} Qe1

{+0.12/19} 41. Qxc4 {+0.00/24} Qxd1+ {+0.11/19} 42. Kb2 {+0.00/25} Qd2+ {+0.09/21} 43. Kb1 {+0.00/24} Qd1+ {+0.03/21} 44. Kb2

{+0.00/26} Qd2+ {+0.09/23} 45. Kb1 {+0.00/27} Qe1+ {+0.09/21} 46. Kb2 {+0.00/22} Qf2+ {+0.09/21} 47. Ka1 {+0.00/24} Qe1+

{+0.03/19} 48. Kb2 {+0.00/80} Qf2+ {+0.09/22} 49. Kc1 {+0.00/20} Qg1+ {+0.09/19} 50. Kb2 {+0.00/23} Qg2+ {+0.09/20} 51. Kc3

{+0.30/14} Qxh3+ {+0.09/19} 52. Kb4 {+0.50/15} Qd7 {+0.09/20} 53. Qc6 {+0.46/15} Qa7 {+0.09/19} 54. a4 {+0.54/16} Qb8+

{+0.09/19} 55. Kc5 {+1.01/16} Qxf4 {+0.15/19} 56. Qe8+ {+0.86/15} Kg7 {+0.09/18} 57. Qxe7 {+0.68/18} Qxa4 {+0.09/19} 58. Qe5+

{+0.68/17} Kf8 {+0.09/20} 59. Qb8+ {+1.35/17} Kg7 {+0.09/18} 60. Qb2+ {+1.27/17} Kg8 {+0.15/19} 61. d6 {+3.13/17} Qd7

{+0.24/18} 62. Qb5 {+4.52/20} Qa7+ {+0.89/18} 63. Kc6 {+4.61/21} Qa8+ {-1.15/18} 64. Kc7 {+4.72/23} Qa7+ {-1.22/17} 65. Qb7

{+4.10/24} Qa5+ {-1.16/17} 66. Qb6 {+4.19/24} Qe5 {-1.20/18} 67. Kb7 {+6.50/24} Qd5+ {-1.82/19} 68. Qc6 {+8.60/18} Qb3+ {-

5.70/15} 69. Kc7 {+9.35/20} Qg3 {-5.91/16} 70. Kc8 {+9.35/19} Qf4 {-3.80/13} 71. d7 {+11.33/19} Qg4 {-3.80/12} 72. Qc3

{+11.94/19} Qf5 {-7.31/13} 73. Kc7 {+11.94/18} Qf4+ {-7.31/12} 74. Kb7 {+11.94/18} Qxg5 {-9.00/14}

descriptiona deep extrapolation about stockfish's top-notch evaluation ( EmptyRe: a deep extrapolation about stockfish's top-notch evaluation (

more_horiz
Permissions in this forum:
You cannot reply to topics in this forum