🚀 go-pugleaf

James Hague <exuhag@exu.ericsson.se> writes:
>Jerry Shekhel writes:
>>
>>I got away with a really fast and really dumb technique in my Chomp game.
>
>"Dumb" is the way to go.  To connect this with the computer-controller
>players thread, I find all the talk about genetic algorithms and neural
>networks to be a major overestimation of what is necessary to create
>a formidable opponent (for action games anyway).

But wouldn't it be interesting to have a computer player that only made
decisions as often as a human player?  (Say, 5 times a second or so.)
Also, regardless of how easy it is to make an overly difficult game, it's
still hard to make computer opponents that aren't very predictable.  If
you play any current Pac-Man game long enough you'll get very good
at anticipating the ghosts' "decisions," because when it comes down to
it they're all canned.  Throwing in some randomness doesn't help a
whole lot either; a ghost whose movements are random enough to be
unanticipateable is usually (in my experience) too "dumb" to be a
threat in practice.

>Just the fact that a computer is making accurate decisions 20-60 times
>a second gives a human player a disadvantage.

This may even be an underestimation (in general).  In my Pac-Man game
the ghosts get to make a decision any time they move a pixel.  They
typically move between 2 and 7 pixels per turn, so that's up to a
couple hundred decisions a second.  (Gee, I never thought about that
'til now.)

>The computer will always know which of two paths is shorter, for
>example, while to the player they both look "about three inches long."

Yes, but what would be more interesting would be a computer opponent
that looked at the board and thought to itself, "I dunno, both those
paths look about three inches long.  But I bet he's trying to snag that
pill over there, so I'll turn left."  Perhaps that's part of the recent
fondness for GA's and NN's in this group.

When you think about it, video games are good test beds for generalized
thinking algorithms (or at least the current approximations that
exist).  One of the video game designer's main tasks is to create a
game world with its own physics.  These rules can be made simple enough
to allow GA's and similar approaches to (potentially) have reasonable
results where similar real-world problems are much, much harder.
(Compare making a robot that follows the white line down the middle of
a highway to an intelligent Pac-Man that motors about the maze.
Pac-Man's sensory input is much more limited and easier to connect to
the "thinking" software.)

Dave Baggett
--
dmb@ai.mit.edu           MIT Artificial Intelligence Laboratory                  
ADVENTIONS: interactive fiction (text adventures) for the 90's!
dmb@ai.mit.edu *** Compu$erve: 76440,2671 *** GEnie: ADVENTIONS