From: Jeroen Tiggelman <Tiggelman <at> CONSUL <dot> NL>
Date: 11 dec 1996
Subject: Re: Evaluation (was Re: Computer Shogi)

>michael facer <michaelf <at> LAVA <dot> NET> wrote:
>>Why is a promoted pawn worth less than a gold? In fact a tokin is worth
>>more. When a gold is exchanged for a gold, it might in certain situations be
>>considered an even trade. (but more likely, the exchange would be to one
>>sides advantage) But when a tokin is exchanged for a gold, it is no even
>>trade. Mr. Aono says that a tokin is worth two generals--in the tokin -gold
>>exchange,you lose a gold, and I gain a gold--that is a difference of two
>>golds (and the pawn that you gain counts as nothing)

Albrecht Heeffer <al @ NETPOINT.BE> wrote:
>This sounds a bit as a contradiction. I would use the same
>argument to say the opposite: A tokin is worth less than a
>gold because you are loosing material by trading a gold
>for a tokin. Even trading a silver for a tokin is bad trade
>unless it is the only way not to loose a gold. The real
>value of a tokin would of course depend on the situation
>on the board. The tokin could be the necessary condition
>for a tsume or could be of no real value at all. If you use
>one static value for a tokin I would make it even to a lance.

Well well, that are two opposed opinions. I really agree with
neither, so we can have a good fight. :-)

In my opinion what we are looking for is not a "value" of a piece
per se, but a value "in context". I am not sure if I made myself
clear before, but I think we should separate counting for the
current position (including the "foothold for a tsume" aspect)
and counting for the value gained in a capture.

For the former count, all that matters it what is on the board.
Therefore, the value of Tokin and Gold should IMO be
considered _equal_.

For the latter count, all that matters is what is gained, therefore
a Tokin should be considered equal to a Pawn.

The evaluation function should probably take both these
counts into account, but I don't think it should really mix them.
There will probably be modifiers, like: my opponent doesn't
have much in the way of footholds, I do, so it's more important
to me to gang up on his King than losing material right now.
And if this is a wild (end) game material m ay again be
considered less important.

Aono's comment interesting, in that a promotion might possibly
be considered to have a value which does not equal the
amount of points gained. However, I plead that it is only valid
if the Tokin may reasonably be exchanged for a Gold in time.
Therefore, we are pointed to a horizon effect consideration:
when we see an evaluated end position where a promotion
just occurred, we really should look a bit further ahead at that
place in the ply tree to see if such an exchange is possible.

This is comparable to, say, the "Pawn-can-run"-modifier in
HiTech (in a chess end game, used to indicate whether the
opponent's king can still get into a pawn's square). Hans
Berliner told in a lecture that adding that feature payed off
very well indeed. (Of course we're talking about a chess
computer with special hardware and geared for pattern
recognition, but...)

Jeroen

P.S. I studied mathematics and computer science,
         in case you were wondering, including some AI. :-)

P.P.S. I will turn to Michael's other comments in another
             mail shortly. :-)