Chess for three gives the White Knight a winning gambit
By ALEX BELL
ALEX BELL is the author of "Games Playing with Computers"
and "The Machine Plays Chess?" Here he explains how
you can test a chess program with some of the methods he has used
over the last 20 years.
DID you know that there are 14,772,512 ways of placing 16 queens
on a 16 by 16 chess board so that no queen attacks any other?
Or that you can mate a black king on an infinite board with a
white king, rook, bishop and knight?
Or that you can be a rotten chess player (like me) and still
test and evaluate chess programs better than Bobby Fischer?
I was asked to comment on three BBC chess programs — Acornsoft
Chess (£9.95), BBC Soft White Knight (£9.95) and Computer
Concepts Chess (£8.95).
If you already have a chess program that allows you to set up
positions then you can compare it with these (and other) programs
using the following standard tests.
TEST 1: MATE TIMER
Ask the program to find two and three move mates from the opening
position. The results were (in seconds):
| |
2 move
|
3 move
|
|
BBC Soft
|
<1
|
7
|
|
Acornsoft
|
9
|
257
|
|
Concepts
|
36
|
>600
|
BBC Soft was by far the fastest to report that there is no solution.
Computer Concepts does not have a specific "mate in N"
mode and has to be set to look at all moves at all levels down
to three and five ply (a ply is a move by black or white). It
chose P-K4 for a full three ply search and I got tired of waiting
for the five ply response.
Acornsoft was intermediate but later testing showed that it
does consider underpromotion which BBC Soft does not.
Testing the programs with genuine two move mate problems gave
an average of about two seconds for BBC Soft, 50 seconds for Acornsoft
and interminable for Concepts.
I was impressed by BBC Soft's speed and compared its performance
using a problem that I have run on many other computers. The oldest
machine was an ICL Atlas (circa 1963 and costing £3 million)
with a program written in machine code which took four seconds
— twice as long as a £10 program on a £400 micro.
But the cycle times of the two machines are comparable — each
could/can add about half a million numbers a second.
TEST 2: OPENING TRAPS
I always play black because there is more opportunity to lay
traps and none of the programs have a book to protect them from
errors. All tests gave the programs about one minute on average
to play their moves. The first game was against the Acornsoft
which, without any forcing, played:
1. P-K4 P-K4
2. N-KB3 N-QB3
3. B-B4
A Giuoco Piano and black has some nice traps, for example the
Blackburne shilling game where you offer up a pawn with 3. N-Q5
— I have never seen a program refuse this one unless protected
by a book so (See Photo II):
4. NxP Q-N4
Dumb programs play NxP forking the queen and rook but QxNP threatens
mate in one and white is in deep trouble. Acornsoft played:
5. BxP+ K-K2
6. BxN QxNP
7. N-B6+ KPxN
8. R-B1 QxKP +
9. Q-K2 QxQ+ +
Move seven is the classic "horizon effect" - anything
to put off the evil check and checkmate. The best line at move
six is to castle (and lose the knight - BQxN) followed by P-KB4.
To test BBC Soft I forced the first three moves and it played:
4. NxP Q-N4
as expected. However one of the nice features of BBC Soft is
that it displays the "Best line" - what it thinks is
its best move and what the opponent will probably reply down to
four or five ply. It clearly saw that QxNP was going to be trouble
but it played the same line as Acornsoft:
5. BxP+ K-K2
6. BxN QxNP
7. P-Q3 KxR +
An unusual move to sacrifice the rook but it does give the black
king a bolt hole.
8. K-Q2 QxQ
9. KxQ RxB
and it's still playing, whereas Computer Concepts was just dumb:
5. NxBP QxNP
6. R-B1 QxKP +
7. B-K2 N-B3 + +
Another trap which can be tried from white opening Giuoco Piano
is the Wilkes-Barre line, thus:
1. P-K4 P-K4
2. N-KB3 N-QB3
3. B-B4 N-B3
which invites white to attack the weak KBP with N-N5. In fact,
none of the programs would play this move (although it is all
right if you know what you are doing) so I forced it on all of
them. Concepts played:
4. N-N5 B B4
5. NxBP BxP +
6. KxB NxP +
7. K-B3? Q-B3 +
8. KxN P-Q4 +
9. BxP Q-B5 +
10. K-Q3 Q-Q5 +
11. K-K2 B-N5 +
12. K-K1 BxQ
BBC Soft also played 5. NxBP inevitable from move six when the
king moved out towards the middle of the board.
White Knight also played 5. NxBP but the B x P+ made it think
for over five minutes. It can do this because it only has to average
one minute for each move (whereas Concepts sticks rigidly to its
time for each move).
The "Best line" showed that it considered K-B1 for
over half of the five minutes because the knight check is difficult
to analyse. However greed overcame caution:
6. KxB NxP +
7. K-K3 Q B3
This is better than Concepts but still greedy - if white retires
the king it cannot have the black rook because of Q-B3, which
either checks or threatens mate in one. But the best line is K-N1
because of the skewer or the following:
8. NxR Q N4 +
9. KxN PQ4 +
10. KxP N-N5 +
11. K-B5 Q-K2 +
12. K-N5 B-Q2 +
13. K-R5 Q-B4 + +
Acornsoft was the only program to avoid the trap with the right
move:
6. K-N1 Q-K2
7. NxR
and black has a poor game.
TEST 3: END GAME PROBLEMS
Comparison of end game abilities is difficult because the time
settings on the programs differ. Concepts always sticks rigidly
to its one minute for each move whereas Acornsoft has playing
levels which have to be set higher for an end game to play at
the test speed.
BBC Soft only averages its moves to one a minute giving it a
definite advantage in difficult positions.
All the programs appeared capable of performing the king, queen
versus king ending (Acornsoft needs level six or seven for this
and announces its move with a chime) but the king, rook versus
king ending was too difficult for all of them.
They all tend to move the king to the centre of the board but
then have no idea what to do with it.
An interesting question is how do these micro-programs compare
with the tournament programs running on large mainframe computers?
Consider the position in Photo I, a famous problem which has
baffled many people since the turn of the century and a good test
of a program's appreciation of repetition of position.
Mainframe programs circa 1975 were just reaching the point where
they could solve such problems because they could be set to a
"play and win" mode, appreciating that repetition is
a draw and underpromoting the pawns.
This is the way it worked. The pawn is given a value equal to
a queen — a fiddle but nevertheless equivalent to what a human
expert does when asked to solve the problem.
The only move that can keep this valuable pawn is P-B7 so the
program should play it without any further investigation of the
consequences.
To misquote Sherlock Holmes: "When all possibilities except
one are eliminated then the exception must be the answer".
Black must now check the white king, and white, in order to
avoid draw by repetition and avoid losing the pawn with a skewer,
must move steadily down the knight file until it can play K-QB2
which prevents the skewer threat.
Black might then play R-Q5 which threatens stalemate (if white
queens the pawn then R-QB5+) but tournament programs in 1975 could
search 11-ply ahead and see that the underpromotion to a rook
is a win for white.
None of the micro-programs (playing black or white) had any
idea of how to play this ending although BBC Soft does have a
crude appreciation of repetition of position and Acornsoft can
underpromote.
I give this example in order to justify my opinion that these
programs' playing abilities are roughly equivalent to the state
of the art about 10 years ago on the most powerful and expensive
computers.
They are incredibly cheap in comparison and the modern micro
is about 10 per cent of the speed of the giants of 10 years ago.
But they are obviously still limited, mainly by their store
size (hence the lack of book openings and knowledge of how to
play end games).
The store size will increase on the BBC Micro when the Tube
arrives. This means that the display handling (which occupies
about 10k in BBC Soft-about one third of the machine) can be done
by the second processor and the programs should then improve their
openings and end games dramatically.
However, it is invidious to compare the modern micro with its
giant ancestors on purely chess playing ability.
The micro, with its interactive features and display monitor,
can and does support vastly superior display and manipulation
features over the old programs.
Ease of setting up positions and selecting options are important
factors in designing micro chess programs and the display of information
is also very important.
For example, BBC Soft display of the "Best line" is
an inspiration by its author, 23-year-old Martin Bryant, and a
feature that will become a must for future chess programs.
To actually see what the program is thinking about while you
wait for its move is quite fascinating, particularly when testing
it with traps.
Another nice feature of BBC Soft is the simple setting of its
playing speed to an average time for each move. When in trouble
the program reacts in an intelligent fashion by taking longer
to make its reply.
CONCLUSIONS
Concepts compared poorly with the other programs in almost every
test and option. In terms of playing strength Acornsoft and BBC
Soft are about equal, with Acornsoft rather easier to manipulate
and with rather more options. It (and Concepts) allows you to
save and reload a position with a filing system, an option BBC
Soft's does not provide.
Nevertheless BBC Soft's "Best line" together with
its very fast problem solving mode must give it the edge in terms
of entertainment and value for money.
Photo I: Acornsoft
Chess
Photo II: BBC
Soft White Knight
Photo III: Computer
Concepts Chess
