Hi, I've been following this thread with great interest.
Thank you all for all amazing information and research in this thread.

I just want to share some information I have gathered on the "PSH" topic.
(Maybe there already is a Super Conny thread about this, but I've not read that, sorry.)

From Kittinger: https://www.talkchess.com/forum3/vie...opic_view=flat:

"Had to research PSH as I suffer from TMA (too many acronyms). 'pre search heuristics' involved recognizing certain specific characteristics and feeding bonus/penalties
into specific moves in the root position move list (ply one move list).
Most famous example of course is the bxh7 sac, although as I recall there were a couple for center pawn sacs when danger of getting constricted to death.
These rules were done from white POV then list run through and swapped for black POV.
With limited processing and search depth, at the time it seemed necessary to look for some typical chess themes cheaply
(applied only for the plyone moves) as opposed to testing lots of positional features deeper in the search.
Obviously there are all kinds of negatives to this approach, but in practice worked out reasonably enough.

As I recall, another short cut which led to interesting play at times had to do with long strings of checks which were not repeats.
With Super Conny at least, these were given a bonus as this was before all the more intelligent extensions for checks had been worked out
(and I think ram space only allowed something like 18 ply maxdepth).
Idea was that such strings of checks might be leading to forced mate or gain of material which program did not have depth to see actual end of."

"Quite a few of my programs had both ply 1 move bonuses/penalties and PVT bonus/penalties.
I spent a lot of time trying to move the PVT values deeper into the search w/out much success.
Never tried that with the PSH though, might work, could have then played to actually achieve or avoid the sac like positions.
The bxh7+ effectively added 5 or 6 ply of depth.
As I recall, the rules for that one required (from white POV) wp on e5, wn access to g5, wq access to h5 or h4. bp at e6 and g7 and no bn access to f6.
Something like that, it was pretty specific rules to give sac high chance of succeeding.
At the time, none of the micro chess programs had a prayer of seeing deep enough to play the sac, even though fairly weak humans could find it.
However, with one-reply to check and singular extensions, I think the sac can be found at much shallower depths, making the whole PSH less valuable."

From Chess Life #11 1984, page 26-27+82: https://uscf1-nyc1.aodhosting.com/CL...84/1984_11.pdf

"Kittinger used a unique approach. He
deliberately simplified his tree-search
evaluation down to an absolute minimum,
coding it very tightly so that it would execute
quicldy. Although this evaluator is not very
smart positionally, it is excellent tactically,
and fast enough to carry out an extended
quiescence search.
To give his program positional knowledge,
Kittinger added an initial positional evalua-
tion routine that examines the board position
before going into the tree search. The pro-
gram assigns "strategy values" for each of
the legal moves, depending on how much
sense they make positionally. It also
values of the pieces depending on posi-
tional considerations, and assesses the
values of the squares pieces are placed on.
This initial evaluation is quite elaborate com-
pared with the evaluation done at the tree-
search level by any microcomputer pro-
gram, and it takes much longer to execute.
But this does not slow the program down,
because the initial evaluation is executed
only once, at the beginning of the search,
while other programs must do positional
evaluations many thousands of times.
Later in 1983, Kittinger extended the Con-
stellation program, adding many more posi-
tional factors to the initial evaluation
routine, endgame improvements, and a
greatly enlarged opening book.

Opening Book. The Super Constellation's
opening book contains a remarkable 21,700
moves, about one-fifth of the lines appearing
in boldface type in the Encyclopedia of Chess
Openings Although this standard opening
book (some variations are 22 moves deep) is
more than enough for most players, the pro-
gram's author has recognized the special
needs of the strong player with an interest in
opening theory. The machine accordingly
features a special learner/trainer mode,
which allows you to "customize" its opening
book with up to 1,200 additional moves.

During the game, both
machines will show you at what depth the
program is searching. For the Super Con-
stellation, that can be up to 20 or 21 plies in
infinite modes; at tournament levels, it
searches to five or six plies in the middle-
game, 10 or more in the endgame.

HIGH-POWERED FEATURES
The Super Constellation s special playing
features illustrate what the future holds in,
chess computers. For example, how many
human players can mate with Bishop and
Knight? This machine can carry out that
mate — as well as all the other basic mates-
at five seconds per move!
Super Constellation's play in pawn end-
games is also quite impressive by computer
standards. To begin with, the program
understands the "square of the pawn" rule,
with such refinements as whether a King
can help its own pawn by taking squares
away from the opposing King. This helps the
program to accurately evaluate threats
created by passed, or potentially passed,
pawns. Super Constellation also knows to
move Rooks behind passed pawns and to
blockade passed pawns.
An unusual and impressive feature is the
machine's ability to sacrifice a piece for a
pawn to reach a drawn endgame.
In the middlegame, Super Constellation
plays an active, tactically oriented game. It'
unique initial evaluation routine also
enables the program to formulate and carry
out positional plans. This does not work
perfectly (neither do my positional plans),
but it greatly strengthens the program's play
and also makes it an interesting opponent
with a relatively "human" style. This initial
evaluation routine contains about 160
specific instructions, which is quite an ac-
cumulation of chess lore for a computer.
Half of these instructions relate to pawn
play, half to play with other pieces. The total
chess knowledge contained in this routine
vastly outweighs anything that could be put
into an evaluator function that must look at
each end position in the search tree.
About 25 percent of these instructions act
negatively; that is, they prevent the machine
from making typical silly "computer —.
moves." For example, many programs tend
to play a move such as ... Bb4 + , "thinking"
that the only reasonable reply is Nc3. Some
programs will even do this with a piece "en
prise, losing after the unexpected reply
c2-c3. Thanks to its initial positional evalua-
tion, Super Constellation avoids this and
similar blunders.
The other 75 percent of the instructions
drive the machine into a search for active
play. Slow maneuvering games are still |
beyond any computer's abilities, but Super
Constellation has some impressive capabili-
ties in more active positional operations. For
example, it knows how to play a minority at-
tack in the Exchange variation of the
Queen's Gambit Declined, and it seeks
pawn breaks to open up a position.
A really impressive feature of the pro-
grams middlegame play is its ability to make
sacrifices on positional grounds. These are
not always perfectly sound (neither are my
sacrifices), but they usually are, and they
nearly always lead to difficult and in-
teresting tactical play. The program knows
how to make the stock sacrifices on e6 in the
Sicilian, and on h7 in the French when Black
lacks a Knight at f6. And it knows how to
follow through on them. It also recognizes
opportunities for positional pawn and Ex-
change sacrifices by either side, and em-
phasizes fast development in openings —
when "out of book" the program will give
back a gambit pawn for development."

/Lars

Hi,
thank you kamoj for all this information, retrieved and shared.

Until quite recently I had not realized the PSH was the only positional evaluation featured in the Super Constellation, this changed with my being able to observe the score behavior; and of course is confirmed by the above information you shared. It is quite an extreme solution; that nevertheless worked well by that time, especially in fast play or blitz.

I completed the Khmelnitsky test, the overall score of the Super Constellation is 1506 (using my usual conditions = 3 minutes thinking time on a position, or 1 minute thinking time on a specific move if needed).





For comparison purpose, the same score has been achieved by Rebel 5 and Saitek Turbo King II D+, whilst the Constellation Expert 5Mhz achieved 1685.

Here is a comparison between both Super & Expert (the Expert features a refined positional evaluation, most probably the root PSH plus later refinements identified deeper in the search tree):


I do think the peak strength in the Strategy domain is a common outcome from the PSH. Anyway, for the Super Constellation, it cannot result from anything else

Despite faster clocking, the Expert does not outperform the Super Constellation in the calculations area; this looks consistent with some additional positional evaluation effort performed along the search. On another hand, the gain in middlegame and opening is quite impressive for the Constellation Expert.

Enjoy the graphs!
Cheers,
Tibono

Miniaturansicht angehängter Grafiken

     

Very very interesting!
Thank you for sharing your enormous amount of knowledge and work on these investigations and graphs!

Ulf Rathsman's programs use the same type of heuristics, but change evaluation for each move/ply like the Expert and later.
Maybe that's why the MM II outperforms the SC as I see in an ongoing thread in this forum, despite MM II's worse handling of pawn structures?
(https://www.schachcomputer.info/foru...ead.php?t=6837)
(Pawn structures were not treated good until the last Rathsman, the Plymate Victoria.)

Maybe Schröder did the same!?
You can browse his algorithms at https://web.archive.org/web/20120320...l/prostuff.htm

Feels like a pawn structure test in the "Khmelnitsky test" suite would have been interesting! ...

Thats a huge difference between superconny and expert, isnt it ?

__________________
Die ganze Welt des Computerschachs

Indeed a huge difference, Thorsten. I would have thought they were closer as far as the software is concerned; just wood, some enhancements and more speed for the Expert... But the evidence is the evaluation management evolved a lot from 1984 to 1985. And surely did not freeze later on, that's why it is a struggle to identify the legacy "PSH effect" in the next Kittinger devices.
My conclusion: the Super Constellation is one of its own, a singularity blessed with some flash of genius!

Hi, I'm not an expert in this, but if you take a closer look at the games played by Super Nova in my "Welcome"-thread, I strongly have the impression that Dave Kittinger also implemented Pre Scan (Search) Heuristics in this little middle class computer!? Do you have any information about this?

Best regards
Roberto

My guess is Dave Kittinger did not jettison the PSH concept, but the more the positional evaluation evolved, the more the effect became diluted.

Here is a comparison graph, including both Beluga and Super Nova:


Of course there are some family genes (this is true with most program series from any programmer), but more specifically I think the peak in Strategy skills is mostly inherited from the inital bare PSH.

6 years later (compared to Beluga & SN), there is no more such a visible footprint from the PSH in the Emerald Classic' Khmelnitsky test (check the Strategy skill):


MfG,
Tibono

Miniaturansicht angehängter Grafiken