From: empath@umcc.umich.edu (Tim Triemstra)
Subject: Game AI posts:
To: jon@stekt.oulu.fi
Date: Mon, 12 Dec 1994 03:18:15 -0500 (EST)

<re-cap of Tim Triemstra's postings from rec.games.programmers>
<with respect to "situation objects" in designing an AI system>
<re-written for general availability on December 11, 1994>

<Tim can be reached at empath@umcc.umich.edu>
================================================================

A method for handling computer intelligence that I'm using is a
little simpler to implement than neural-nets: not necessarily
better or worse, maybe just "different."

All units in any type of strategy game have limitted situations
that they can be put into (the broader the definition, the
smaller the number.)  IE: If you have a game with 10 tanks on
each side it could be said that there are only 11 possible
situations: 1 situation for each number of tanks visible to the
player and 1 situation where none is visible.  Of course, that
is too broad to really _do_ anything with.  To be of any real
value the situations must take into account strength, ranges
and other aspects of the situation.

Since I have a smaller number of units in my game, I make little
"situation objects" composed of all the various possible
situations.  Each situation object also contains the orders to
give to the player.  The orders are executed if it is determined
that this particular "situation object" is the best match for
the current state of the player.  The higher the number of
possibilites, the greater the intellect of the computer
controlled player.

To me, this was much simpler than hand-coding a tree because I
could make a move construction kit that builds these situations
and the resultant decision and saves them into a library.  Beta
testers can help build them too.

<... end of message #1 ...>

> Wait, so you build a (let's call it) a data base for each
> instance of tank vs other tanks and then a "good" solution for
> it???  Wouldn't that be huge?

Yes, in a SSI wargame it would not be a viable solution.
However, in my game (and I'm sure many other people's) there are
limitted types of units and limited quantities of units in each
mission or scenerio.

In my game, there are only 5 units (I may merge them down to 3
total types of units) which can have various attributes (larger
weapons, higher speeds etc.) but those factors are taken into
each database, or "situation object" entry.

In my early designs, only 30 or so data elements is needed for
each unit type to give what appears to be nearly adequate
responses.  I also curb the results slightly by giving units
commands like "move here stealthy" or "move here while attacking
whomever you see" to stack the deck in the decision making.

Mind you, this idea of "situation objects" is purely one way to
do it that may be practicle in some gaming situations.  I'm not
a big fan of simulations (like SSI games) and hence wouldn't do
a game like that.  Those SSI games couldn't use my idea, but
Dune II could (and it would actually be smarter :)

<... end of message #2 ...>

Welp, the idea of finding the nearest "situation object," as I
sometimes refer to it, isn't nearly as difficult as you may
think.

First off, these objects are based completely on "relative
situations"  ie: if a guy is stronger and to the upper-left
of the character then that would be one situation.  So, the
tree goes something like this:

1) How many opponents are visible?
 ( branch to the decision objects sorted for each answer )
 - branch if 1 or more
2) How many can hit me within X moves
 - branch if safe to not run
3) How many can I hurt?
 - branch if any
4) _SHOULD_ I try to hit him or run
 - branch if try to hit
5) Which is most benefical to hit
 - branch etc...

Now, in my particular game, each unit is also given priorities
that factor into this decision.  So, you may tell a unit to
scout but give him the order to engage or hide, this of course
adds weight to different tree decisions.

The major advantage of a system like this, as I said, is that
you can make the AI as good or bad as you like.  Quite frankly,
I'm under the opinion that a "game" does not need particularly
good AI to be a good game.  Simulations (not my favorite) are a
different story.  A game need only provide an enemy adequate
enough to give the player some fun and to enact the plot.
Dune II (my favorite example) has horrible AI, and anyone can
eventually beat the game, but it is unquestionably enjoyable to
play.


-- 
Tim Triemstra ... Empath Software .... Empath@umcc.umich.edu
<><><><> If'n you ain't the granddaddy of all liars <><><><>