| |
Main
Date: 04 Jan 2009 00:44:29
From: Sanny
Subject: GetClub game improved. (2 new improvements done)
|
Last 2 weeks there was a huge modification but later it was found that
it actually slowed and the improvements were in wrong direction. So
new modifications were not made.
However 2 bugs were found in the process and removed.
1. Now I hope GetClub will always play the opening moves. Lets see new
games and see if it always plays the openings.
2. 5 new Openings were taught to GetClub as Help Bot was able to play
many new openings and then GetClub had to think a lot. Now with new
opening learnt it will play good moves in return to your new openings.
If you want to enjoy Chess play a few games with Beginner Level and
enjoy.
Play Chess at: http://www.GetClub.com/Chess.html
I hope now the playing exprience has improved a lot.
It will play quite strong moves as the bugs have been removed. It
would be very challenging to win against the Beginner Level. It will
not let you win easily.
Bye
Sanny
Play Chess at: http://www.GetClub.com/Chess.html
|
|
| |
Date: 05 Jan 2009 20:51:56
From: help bot
Subject: Re: GetClub game improved. (2 new improvements done)
|
On Jan 5, 4:33=A0am, Sanny <[email protected] > wrote:
> > =A0 With GetClub, it sometimes happens that we
> > are chugging along, and POW!, the program
> > will sacrifice a piece for mere pawns and
> > maybe a spite check; game over.
> If this happens again inform me with the link to the game and I will
> see if positional values need to be decreased.
>
> At the moment GetClub give +1 point penalty for any opponent piece
> reaching near the King. So that king is kept away from attacks.
>
> Should that be changed?
Yes.
For instance, in simple endings it often happens
that a player should use his King to restrict the
enemy King-- not stay on the edge of the board
or worse, near a corner. In this case, it is illegal
for one King to move right next to the other, but
in many other cases, the close approach of an
enemy piece is not necessarily dangerous.
Let's say you have a position where GC can, if
it wants, sacrifice a Knight for two pawns and a
spite-check which brings, for a moment, one of
its pieces next to the enemy King-- which is
perfectly safe from being checkmated. In this
case, the Knight-sac may well lose, and yet
the opponent has done nothing to earn victory,
except gobble up the unsound sacrifice and
then move out of check.
A far better way to promote successful attacks
is to let the program try to find mate or the win
of decisive material via its normal calculations.
Value the King at 99 points or more, so that the
program never, ever "sacrifices" its King, for
compensation.
In reality, a spite-check has /no value/ apart
from annoying the opponent (hmmm) and if it
might be followed by a number of losing replies,
but only one, difficult-to-discern, correct defense.
In general, computers do well to sit back and
wait for tactical errors by their human opponents,
then POUNCE. This is followed by the most
annoying thing of all: near-perfect technique in
mopping up a won game! Coming back from a
dead lost position generally requires some
tactical error or errors-- often ones which most
computers are very unlikely to make.
Some programs are far more aggressive than
others, but I seriously doubt that any really
strong ones score an enemy piece incursion, in
itself, so high as equivalent to a full point. That
scoring strongly encourages spite-checks, but
the goal is not to merely inflict spite, but to
wreak havok like Conan the Barbarian. Think
Collateral Damage, not "I spite-check you".
-- help bot
|
| |
Date: 05 Jan 2009 01:33:26
From: Sanny
Subject: Re: GetClub game improved. (2 new improvements done)
|
> =A0 With GetClub, it sometimes happens that we
> are chugging along, and POW!, the program
> will sacrifice a piece for mere pawns and
> maybe a spite check; game over.
If this happens again inform me with the link to the game and I will
see if positional values need to be decreased.
At the moment GetClub give +1 point penalty for any opponent piece
reaching near the King. So that king is kept away from attacks.
Should that be changed?
Bye
Sanny
Play Chess at: http://www.GetClub.com/Chess.html
|
| |
Date: 05 Jan 2009 01:28:26
From: Sanny
Subject: Re: GetClub game improved. (2 new improvements done)
|
> In a previous game it threw away the exchange without any
> compensation...twice!
> The root cause for this sort of thing, I suspect, is too much weight
> on positional aspects in the calculation of the evaluation score.
Yes, there is too much points attached to positional values. It is
done so that king is protected from intruding pieces.
It gave away knight for one pawn because of below benefits.
If Knight had played Nd7
1. It would have blocked the Black Bishop. (1 point)
2. The Center would have been controlled by White pawn at e5. (1
point)
In case it plays Bishop to b4, It gets points for Piece Advancements.
And when Bishop takes the knight. White will get double Pawns. So
again 1 point for that.
So it sacrificed Knight for 1 Pawn + 3 strong Positional advantages.
What do other Chess engines tell about this sacrifice?
Bye
Sanny
Play Chess at: http://www.GetClub.com/Chess.html
|
| | |
Date: 05 Jan 2009 15:50:32
From: David Richerby
Subject: Re: GetClub game improved. (2 new improvements done)
|
Sanny <[email protected] > wrote:
> If Knight had played Nd7
>
> 1. It would have blocked the Black Bishop. (1 point)
The black bishop is already mostly blocked by the pawn on e6. It's
one of the characteristic features of this opening.
> 2. The Center would have been controlled by White pawn at e5. (1
> point)
>
> In case it plays Bishop to b4, It gets points for Piece Advancements.
> And when Bishop takes the knight. White will get double Pawns. So
> again 1 point for that.
>
> So it sacrificed Knight for 1 Pawn + 3 strong Positional advantages.
No. It gave up a knight for a pawn and a few minor positional
advantages.
> What do other Chess engines tell about this sacrifice?
They think it's garbage.
Have a read of the following article by Dan Heisman:
http://www.chesscafe.com/text/heisman57.pdf
Heisman states what he calls the Principle of Tactical Dominance:
Tactical criteria dominate positional criteria. Therefore, use of
positional criteria is almost always useless if there is a tactic
that wins material or checkmates; decide tactics first and only
apply positional criteria if no tactic exists.
Dave.
--
David Richerby Miniature Metal Priest (TM): it's
www.chiark.greenend.org.uk/~davidr/ like a man of the cloth that's made
of steel but you can hold in it
your hand!
|
| |
Date: 04 Jan 2009 12:06:31
From: zzz
Subject: Re: GetClub game improved. (2 new improvements done)
|
On 4 jan, 09:44, Sanny <[email protected] > wrote:
> Last 2 weeks there was a huge modification but later it was found that
> it actually slowed and the improvements were in wrong direction. So
> new modifications were not made.
>
> However 2 bugs were found in the process and removed.
>
> 1. Now I hope GetClub will always play the opening moves. Lets see new
> games and see if it always plays the openings.
>
> 2. 5 new Openings were taught to GetClub as Help Bot was able to play
> many new openings and then GetClub had to think a lot. Now with new
> opening learnt it will play good moves in return to your new openings.
>
> If you want to enjoy Chess play a few games with Beginner Level and
> enjoy.
>
> Play Chess at:http://www.GetClub.com/Chess.html
>
> I hope now the playing exprience has improved a lot.
>
> It will play quite strong moves as the bugs have been removed. It
> would be very challenging to win against the Beginner Level. It will
> not let you win easily.
>
> Bye
> Sanny
>
> Play Chess at:http://www.GetClub.com/Chess.html
Based on a game I played against "Easy" earlier today, there isn't
much reason for optimism on the evolution of GetClub's playing
strength.
I had white, the game went 1. e4 e6 2.d4 d5 3.Nc3 Nf6 4.e5
After 4.e5 GetClub appeared to be out of book, which is surprising
considering this is one of the main lines of the French Defense.
The move it came up with was even more surprising though: 4...Bb4???
gives away a piece for, well, nothing much. Game over after another 19
moves.
In a previous game it threw away the exchange without any
compensation...twice!
The root cause for this sort of thing, I suspect, is too much weight
on positional aspects in the calculation of the evaluation score.
GetClub needs to (re-)learn some respect for material!
Therefore my advice to Sanny is: drastically reduce the weight of the
positional part in the evaluation function.
Let the program rely mostly on material calculation. Don't burden it
so much with misunderstood and overestimated chess "knowledge".
As an experiment: divide the total of the positional terms in the
evaluation function by 10, run your tests and compare.
|
| |
Date: 04 Jan 2009 01:33:34
From: help bot
Subject: Re: GetClub game improved. (2 new improvements done)
|
On Jan 4, 3:44=A0am, Sanny <[email protected] > wrote:
> Last 2 weeks there was a huge modification but later it was found that
> it actually slowed and the improvements were in wrong direction. So
> new modifications were not made.
>
> However 2 bugs were found in the process and removed.
>
> 1. Now I hope GetClub will always play the opening moves. Lets see new
> games and see if it always plays the openings.
>
> 2. 5 new Openings were taught to GetClub as Help Bot was able to play
> many new openings and then GetClub had to think a lot. Now with new
> opening learnt it will play good moves in return to your new openings.
>
> If you want to enjoy Chess play a few games with Beginner Level and
> enjoy.
>
> Play Chess at:http://www.GetClub.com/Chess.html
>
> I hope now the playing exprience has improved a lot.
>
> It will play quite strong moves as the bugs have been removed. It
> would be very challenging to win against the Beginner Level. It will
> not let you win easily.
This sounds about right as at first, I was winning
even easier than before, but more recently, I've had
some difficulties against the Beginner level.
In addition, I've noticed an improvement in the
openings play-- things aren't so easy for me
anymore. I've been taking White a lot and GC
seems to have truly "improved" in such lines as
the Benoni, the Queen's Gambit declined, and so
forth. I may have to switch to playing the Black
side and use my "secret weapons", such as the
Larry Parr line in the QGD (Black comes out on
top), or -- heaven forbid -- just make up new stuff
on the fly, to confound the machine.
These computers can be really daft at times.
While I disagree with the great Dr. IMnes as to
the value of the Ruy Lopez "evidence" of dafticity,
I noted that in analysing game one of the 1972
Spassky/Fischer match, where Mr. Spassky
played p-b6!, even Rybka -- the greatest of the
computer chess breed -- made a beginner-style
blunder in taking on a6, which nets a very
simple draw (because White's a-pawn cannot
Queen).
With GetClub, it sometimes happens that we
are chugging along, and POW!, the program
will sacrifice a piece for mere pawns and
maybe a spite check; game over.
On the other hand, I saw a recent game -- a
brilliancy prize winner, no less -- in which one
of America's top grandmasters thought his
opponent had a perpetual check, when he
didn't. GM Shabalov could have snarfed a
Rook, with check, and according to Rybka,
he then could go ahead and snarf a Bishop
to boot! The checks merely drive his King
to safety... .
-- help bot
|
|
