article1085.asp.htm

电脑游戏中的人工智能制作（含源代码）

controlling it, time for more to get there &c. Thiswould provide an
incentive for the AI to encircle enemy territory.

Robert Uhl

==============================================================================

Robert A. Uhl ([e-mail]) wrote:

:   Heh. Actually, if X is driving through Y, in order to create two
: fronts along which Y must fight, thereby forcing Y to regroup, then
: 'b' has been executed.


    Good point.  Yet another example of how a human can spot an opportunity
in the face of the most daunting situations....


: >    If we define a contact point, then does that give us a natural focus
: >towards which to direct our forces and our strategic 'thinking'?  They would
: >seem to.

:   Yes. The AI should attempt to create as many contatc points as
: possible. Why? These points mean that he is actually fighting the
: enemy and hopefully killing the same. That is why two parallel lines
: are a good formation. With a good lookahead, though, the T formation
: is also good, because it will enable many points of contact _over
: time_, and hurt Y's ability to create them. Part of the battle is
: controlling contact. In fact, if the AI can do this controlling, then
: it is my belief that it will have an extremely good chance at winning.


   I'm not sure I agree that I would want the AI to maximize its number
of contacts with the enemy; I would agree that it should seek to control
them.

   Maximization in itself will only lead to toe-to-toe WWI slugfests, and
basically leads to the AI playing a war of attrition.  That's perhaps one
of the defining characteristics of most AIs today--if they don't cheat
somehow, then they tend to fight wars of attrition.


:   The sphere of influence idea seems to work well with the points of
: contact idea. Perhaps each unit has a sphere wherein it can contact
: within a certain number fo moves (say 3), and its job is to contact
: the fewest enemies at once but the most over time. IOW, it doesn't
: want to be outnumbered but it wants to see action.


    Yes indeed.  That's sort of where we were headed, I believe.
I do like the idea of factoring time into the equation somehow; the AI
ought to be willing to have NO enemy contact for two turns if it's busy
moving forces around for MAJOR enemy contract (say, the Normandy invasion)
on the third turn.  That does make sense intuitively.  Perhaps we can tie
time-weighted values to multi-turn engagement decisions?

:   And two units have a greater sphere of influence, say 8 moves, than
: just one. This would help reflect the greater power of two. Controlled
: territory would be defined as that surrounded by my own units and
: without enemy units, once again utilizing the SoIs. Contested would,
: of course, be that which neither side controls. Perhaps a 'strength'
: value would be attached to areas, indicating the number of units
: controlling it, time for more to get there &c. Thiswould provide an
: incentive for the AI to encircle enemy territory.

   Agreed.  This is very similar to the idea of using a 'fire distribution'
map that was presented by Daniele, actually, just a different method of
solving the same problem.  We'll want to do SOMETHING like this in order
to properly consider the limiting effects of enemy weaponry and the
local terrain.

Steven
==============================================================================

On 15 May 1995, Andrae Muys wrote:

> Steve Woodcock ([e-mail]) wrote:
> : Andrae Muys ([e-mail]) wrote:
> 
> : :             YY      Now by any standards X is in a bad way.  It has been
> : :              Y      completely outflanked and his left flank is already
> : :      XXX     YY     overrun.  Intuitivly his front is now perpendicular
> : : XX XXX  XX   XY  Y  to Y's.  I think we may need a concept such as
> : :    X        X Y     contact point, which in this case is Y's centre, and 
> : :                 Y   X's left flank.  Naturally in most battles there would
> : :             YYY     be multiple contact points.  Personally I would draw the
> : :             Y       fronts as follows.
[snip]
> benifits of creating/avoiding specific contact points.  e.g. in the 
> example above X should avoid contact with Y until he has rotated his 
> front.(it looks like we may still need to consider fronts as well).

Two options I can see here - either X moves its forces into Y's "front" 
to create as much damage as possible (in a concentrated strike or 
"blitzkreig" style attack) or X moves its "front" back, forcing Y to make 
the next move (allowing X the advantage in defence).

Is X aggressive or defensive? Y's forces are rather spread out, so X can 
"spear" through the front and attack from both sides (like an inverse of 
the "pincer" movement).

Just a thought...
-Alex
==============================================================================

I appologise for the excessive quoting but without the diagram any reply 
is awkward.

Satrapa / Alexander Marc (ISE) ([e-mail]) wrote:
: On 15 May 1995, Andrae Muys wrote:

: > Steve Woodcock ([e-mail]) wrote:
: > : Andrae Muys ([e-mail]) wrote:
: > 
: > : :             YY      Now by any standards X is in a bad way.  It has been
: > : :              Y      completely outflanked and his left flank is already
: > : :      XXX     YY     overrun.  Intuitivly his front is now perpendicular
: > : : XX XXX  XX   XY  Y  to Y's.  I think we may need a concept such as
: > : :    X        X Y     contact point, which in this case is Y's centre, and 
: > : :                 Y   X's left flank.  Naturally in most battles there would
: > : :             YYY     be multiple contact points.  Personally I would draw the
: > : :             Y       fronts as follows.
: [snip]
: > benifits of creating/avoiding specific contact points.  e.g. in the 
: > example above X should avoid contact with Y until he has rotated his 
: > front.(it looks like we may still need to consider fronts as well).

: Two options I can see here - either X moves its forces into Y's "front" 
: to create as much damage as possible (in a concentrated strike or 
: "blitzkreig" style attack) or X moves its "front" back, forcing Y to make 
: the next move (allowing X the advantage in defence).

IMO The first of your options is probably going to lead to disaster.  X 
is not only outflanked but overrun.  From this postion any attack he 
makes is going to be piecemeal, and forces commited to an attack 
piecemeal are destroyed piecemeal to very little effect.  IMHO the second 
option is the only option avaliable to X, the question for X is how far 
back to regroup, and what to do next.  Some of his options as I see them 
involve, pulling all units at the contact point back, all units not at 
the contact point foward, form an offensive formation and try to break 
Y's, by now dissordered, centre. Or maybe send forces to delay Y's flank, 
utilise defence in depth at the contact point to buy time, and either 
prepare an attack, prepare a defence, or use the manuvour to start a 
rearguard action.

: Is X aggressive or defensive? Y's forces are rather spread out, so X can 
: "spear" through the front and attack from both sides (like an inverse of 
: the "pincer" movement).
Personally I would quite enjoy gaming Y against an X attacking from this 
postion.

Andrae.
==============================================================================

Andrae Muys ([e-mail]) wrote:

: Of course the ultimate AI wouldn't find itself in such a dangerous 
: position.  A bit like the untimate General wouldn't.  But if it does it 
: must extricate itself well.  Just one more 'special' situation to test 
: the AI's abiltiy.


    War appears to be full of special cases, eh?

    ;)

: :     If we define a contact point, then does that give us a natural focus
: : towards which to direct our forces and our strategic 'thinking'?  They would
: : seem to.

: Well this thread is useful.  The idea of contact points should radically 
: prune any decision tree.(And sooner or later the AI will have to make a 
: choice)  Of course at the stratigic/grand stratigic levels we may need a 
: modified definition of contact point but at the level I am interested in 
: Contact points appear to be a good way to look at things.  In fact now 
: that I think about it, contact points are how **I** allocate **MY** 
: consideration.  This approach however leads us to consider how to 
: recognise potential contact points and how to evaluate the relitive 
: benifits of creating/avoiding specific contact points.  e.g. in the 
: example above X should avoid contact with Y until he has rotated his 
: front.(it looks like we may still need to consider fronts as well).


   Yes, I'd come to that conclusion as well.  By identifying contact
points and making them the focus of (at least some) activity, we've 
pruned our number of decisions fairly radically.

   One concern I have with relying on them exclusively, however, is
that this makes the AI very reactive without some additional logic to
force it to 'seek' the enemy.  Two obvious ways to handle this would be
a.) random movement of forces until contact with the enemy has been
achieved and b.) deliberate movement of forces to attain certain objectives
(which are, in effect, 'artificial' contact points, if you will).

   It would be simple enough (I'm still talking Napoleonic-level for
simplicity's sake) to provide the AI with whatever its objectives are
for a given scenario; that's easy.  The hard part is how to make it
approach those objectives in some fashion that makes sense rather than
to blindly throw its units down a road leading towards said objective
(perhaps the main failure of the AIs in Perfect General and Empire Deluxe,
to name two examples).

   As I recall, Napoleonic-era tacticians were trained to recognize
'classic' battle formations (one such example being the 'T' you presented
earlier) and react accordingly.  Pattern recognition is easy enough
to do on a computer via a variety of methods, ranging from simple table
lookup to full-blown neural nets.
   
: :    If we use your previous suggestion for identifying centers, combined with
: : the above-mentioned contact points, then this may lead us towards a more
: : natural way of handling the above situation.  Based on what we've discussed
: : so far, I would envision an AI's logic train going something like this:

: :         Pass #1:  Identify natural 'groups' of X-Y forces using the
: :         'circle method' discussed earlier, perhaps taking into account
: :         the possibilities of influence and interdiction as a previous
: :         poster suggested. 

: :         Pass #2:  Having identified these natural groupings, identify
: :         contact points amongst the forces within each group.  These
: :         will serve as natural foci for our planning process.

: :         Pass #3:  Having identified natural groupings and focus points,
: :         we now begin thinking about steps needed to link up our groups,
: :         minimize the size of enemy-held areas, elimination of enemy units,
: :         etc.  
: This sounds like it should work reasonably well, however I have a feeling 
: there may be problems with the way it handles many sparse stratigic 
: situations.  I'll think about it and we can discuss it futher when I have 
: clearified my concerns.


   Agreed, it may.  However, based on the above idea of using pattern
recognition to manuever our units in something resembling a smart fashion,
we can now elaborate on Pass #3:

      Pass #3:  Having identified natural groupings and focus points,
      run a pattern-matching algorithm to identify most likely tactical
      situations and reasonable responses.  Randomly pick from the best
      2 or 3 to avoid predicatability.  For each turn a course of
      action is in place, make some evaluation of its effectiveness
      and/or success (i.e., has the pattern changed signficantly enough
      to warrant a new course of action?  have we attained the objective
      in question?  have we lost contact with the enemy?).


   In the case of a sparse strategic situation, the AI defaults towards
trying to attain known and logical objectives (i.e., moving towards Paris).
Once contact is made, the situation naturally becomes less sparse, and
the AI begins to make moves based on sound engagement philosophies.

   The AI will end up being a bit 'bookish', if you will, but certainly
ought to surprise you once in a while.

: However as discussed above at a lower level fronts again become 
: important.  IMHO this is because tactical considerations require physical 
: cohesion, while stratigic(level) utilise logical cohesion.  
<snip>

   I'll buy that.   That's a good definition, particularly for this
problem (Napoleonic-era combat), and better than many I've seen.

: [regarding comment that we post this to rec.games.miniatures]

   Actually, I've already seen one thread on comp.ai recommeding THIS
thread as 'worthwhile'.   ;)



Steven

==============================================================================

In article <3oro0c$[e-mail]>, ccamuys@