The "Lambda Parameter"?

[Guest]

Archived from groups: comp.ai.games (More info?)

Greetings Fellow Newsgroupies,

I work for a video game company producing games that would fall under
the category of "First Person Shooter". These are heavily 3D
simulated environments (forgive me if you are already familiar), where
the game involves using a handful of simulated weapons to kill others
and not be killed.

I often find in writing AI for these games that there is a "sweet
spot" of behavioral complexity where the allies and enemies provide
enough variety to make the game challenging, but not so much as to
appear erratic or indecisive. To be honest, I was quite surprised by
this when I first ran into it.

Starting with AI for such a game is simple enough:
- If you can't see the enemy, move until you can.
- If you can see the enemy, shoot at him / her.

This is an effective strategy, but very simple and predictable,
clearly "dumb" as our target audience would rightly criticize. Now
add a third behavior:

- Occasionally move to a nearby random location.

And suddenly the game becomes infinitely more complex and considerably
less predictable. It's tempting at this point to say "if one extra
behavior is good, two is better". However having explored that road
in painful detail, I have yet to find any additional behaviors that
substantially add to the qualitative appearance of Intelligence, and
in fact further detailed attempts at things like the so called "squad
AI" or "tactics" are more often a detriment. I suspect that a similar
situation holds for most nondeterministic games, including something
like GO.

So where are we going here? In so far as AI playing a game as an
opponent or an ally, is there such a thing as a behavioral complexity
"Lambda Parameter" between too chaotic and too static? And if so, how
does that affect our concept of AI as a game player and AI as a whole?


Just Curious,
Christopher Reed
Raven Software / Activision

 

[Guest]

Archived from groups: comp.ai.games (More info?)

Your point of view is quite interesting and hopefully you'll find
that lambda parameter to progress level of reality of AI opponents.

Though I'd like to say it doesn't exist. Situation you mentioned
is a one on one fighting scenario and is described well,
meaning additional AI would make it more erratic but not smarter.

Compared to human evolution this is equivalent to pre stone age.
Later on humans learned to fight in groups and rises effectivity
in terms of killing - of course. This new tactic is more challenging
and used in Far Cry and pleasures player - you may know.

Acting as a compact group or flock of opponents implys
communication between individuals. So you may say language
is that searched lambda parameter, but it is not a parameter
and not a technology it is kind of sibling of consciousness.

Personaly I wouldn't describe the difference between
grass hoppers and ants as a parameter. If biologists discover
this difference as a switch to turn flies into state building insects,
you're right then.

Don't bother me as somebody detecting mind in ants. As long
as nobody has a provable test to find consciousness in human
brains it is not worth to discuss.

So train your AI opponents to talk to each other, make them
aware of each other, let them know how to fight together, give
them tools to develop an adequate strategy, the ability to forget
a failed one and memory to learn from failures.
Balance maximum kills and survival.

Then one day they will learn to make peace not war.

Regards

 

[Guest]

Archived from groups: comp.ai.games (More info?)

T.Becker wrote:

> So train your AI opponents to talk to each other, make them
> aware of each other, let them know how to fight together, give
> them tools to develop an adequate strategy, the ability to forget
> a failed one and memory to learn from failures.
> Balance maximum kills and survival.


Interesting point. Communication between actors is highly subjective to
noise. E.g. explosions may delay communication, comm-links may be
sabotaged, terrain occlude potential addressees, emotions may get in the
way. Therefore, whether 2 or more actors team up together can be
described in those terms.

Regenerating the exact same sequence for the AI actors would be
extremely hard, because the situation has been defined as a function of
time, player actions and positional location information.

If everything is wired onto that communication model, then the extremes
tell you that either all communication is effective, and you get more or
less predictable behavior, or communication is broken, in which case you
get other predictable behavior. In between, actors will switch from
'group thinking and group controlled' to 'individually thinking', which
is where lambda would be high. In most games, the player brings lambda
down, as his missions are all about killing and disrupting the enemy
communication. The game pumps up that lambda, usually by cheating,
making sure that all 'communications' out of sight of the player are
successful.

Building such a comm layer is another can of beans though..

Cheers,

Ignace Saenen
PlayLogic Games

 

[Guest]

Archived from groups: comp.ai.games (More info?)

IS> Building such a comm layer is another can of beans though..

A tiny one and it would be worthwhile. I don't think forcing developper
to put power in writing sophisticated shader code is an better option to
create more realistic and smarter games.

A comm layer is needed, but doesn't hit the point. Only minimal bandwith is
needed . An implementation of cognition will take next step. Let the AI deal
with information, give them different goals depending on their current
situation and look what happens.

They should know what to do in case of partial information and how to get
more information using more or less intact communication channels. Level of
reality will grow exponential with amount of action alternatives.

Code examples of counterfactual reasoning are available by google, who will
serve that can of beans first?

Regards

 

[Guest]

Archived from groups: comp.ai.games (More info?)

Ignace Saenen <ignace.saenen@NOSPAM.writeme.com> wrote in
news:2l0qk4F790g8U1@uni-berlin.de:

> If everything is wired onto that communication model, then the extremes
> tell you that either all communication is effective, and you get more or
> less predictable behavior, or communication is broken, in which case you
> get other predictable behavior.

Im stepping in late to the conversation but
how about "machine" + "random" = "more human"?

If there is a random chance that a message doesnt make it, then some
interesting errors of decision could be made. Which would in fact create
some funny and realistic situations.

If you give pluses and minuses to the random based on terrain, weather
conditions, actions of the units (units at rest would do better than units
involved in moving actions); then you might get very close to a "human"
situation.
"It is not logical, but you will often find that it is true"

Gandalf Parker
-- Do you wish for Artificial Intelligence?
Or do you wish for Artificially Human?
The two are very different goals.