[Freeciv-Dev] Re: Want calculation (Was:Re: flying AI (PR#1162))

["Ross W. Wetmore" <rwetmore@xxxxxxxxxxxx>]

At 11:13 AM 02/01/01 +0000, Gregory Berkolaiko wrote:
> --- "Ross W. Wetmore" <rwetmore@xxxxxxxxxxxx> wrote: 

>> To repeat a former caution though. There is a rationale for everything
>> if you can hold onto your sanity long enough to figure it out. But
>> these
>> long involved calculations are not in themselves particularly important
>> or useful. They attempt to mimic some human assessment process, but
>> don't
>> really cover the full spectrum, so accuracy is meaningless within a
>> fairly broad range.
>
>I disagree on "mimic some human assessment" issue.  It's an attempt to
>code mathematically sound tactics IMHO.  Which is "to profit whatever we
>do".  The human assessment is more "can I kill it somehow?" while the
>code attempts at "will I profit if I try to kill it?" approach.

I guess it depends on your estimation of human assessment abilities. Good
human players IMHO don't just kill for pleasure :-).

And what the computer assesses is a small fraction of what a good human
player will do. For example there is no consideration of multiple attack
strategies, order of unit movement or placement (a good human may 
completely populate a city radius to block partisan creation, or move
units in an order that allows ZOC-free movement).

Even in single attacks, there is no consideration of potential weakness
generated by unit loss or damage, and possibility of counter attack.

So a really accurate "profit" is unfortunately sadly lacking in a number
of key areas, and faith in mathematically sound as a better way to go
might be very misleading :-).

But even in your profit calculation, it is pretty much a binary choice
that you are interested in, and accuracy is not critical apart from the
flip point (which a human will realize is fuzzy anyway).

>Possibly bigger difference is in general view on combat:
>Human "here is the enemy, where is my Howitzer?" 
>and AI "here is my howitzer, where is the enemy?"

Actually, it is more - 
  computer -> limited scope integral logic, 
  human -> broad scope fuzzy logic.

 ... which is why computer programs less than BigBlue aren't generally 
good opponents :-).

The point I'm trying to make is not that accurate assessment of something
is bad, but that excessive concern for accuracy in a few limited cases 
over lots of cruder assessments of a wide variety of factors will be
counterproductive. And in the end, there is no single "best" move or
action, just a fuzzy range of good and bad things. The AI goal should
be to always pick something good, with no really bad side-effects, plus
a certain amount of variability in the choice == unpredictability is
much more effective (and interesting) as a game strategy against humans.

>Best,
>G.
>__________________________________________________
>Do You Yahoo!?
>Everything you'll ever need on one web page
>from News and Sport to Email and Music Charts
>http://uk.my.yahoo.com

Cheers,
RossW
=====