Adaptive
AI?
Was thinking about AI last night. So without further adieu, some
ideas:
AI
capabilities.
Let's take a Warzone AI for an example. Basically a WZ AI is 90%
random, the other 10% covering general stuff that needs to be done
like checking that it has all required buildings in the Required
Buildings list (HQ, at least one factory etc). Warzone's AI is quite
simplistic really, at least in comparison with later games, but it
gets the job done. However, it get's it done by random chance, which
is something I'd like to see change.
Here's a brief rundown of what a Warzone AI does throughout the course
of a game, with notes on what should be happening instead (going from
the pre-game loading sequence, I'll go into some thoughts about the
pre-game menu another time):
*1 Player clicks ok to start the skirmish. Pre-game loading sequence
initializes. At this point there are some things I want to see
happening, I'll talk about that in a moment.
*2 Skirmish Starts. AI starts methodically checking it's triggers to
see what it needs to do, including assigning it's trucks to building
duties based on the Vital Structures list. Vital Structs = Oil, power,
HQ, Factory and so on. As soon as first Factory is built, AI will
randomly pick templates for units and start building, with trucks
being build sporadically throughout.
*3 Continue checking triggers. Send out scouts (usually trucks for a
start) to check for oil resources and cap them, and also scouting out
the general area. The AI knows at all times exactly where the other
AI's are as well as the human player (s).
*4 Check Triggers. Once the AI has built all vital structures it will
start to concentrate more heavily on base defence as well as building
additional structs such as factories. Also keeps track of research
topics (again, randomly assigned). Once vtol tech has been reached it
will build a vtol fac and proceed to start constructing vtols on a
triggered basis.
*5 There are 3 triggers the AI is primarily interested in regards
attacking - AttackStuffTr, and allOutAttackTr are the battle oriented
ones. The first trigger will be set to a lower number such as 300,
which means that every 30 seconds (10 WZ units = 1 second) the AI will
harass another player with a small group of units. The second trigger
will typically be set to 1500 or more and when triggered the AI will
attack with a massed force. The other trigger (numAttackers)
determines who the AI will attack. This is normally set to around 40,
meaning that if the AI has less than 40 units it will harass an AI, if
it has more than 40 units it will cause big trouble in the Player
Base.
*6 Continue checking triggers, keep building base items and defences
while attacking AI or player. As structures are destroyed, send out
trucks to the dumbest locations possible to try and rebuild (bit of
sarcasm there). If the enemy attacks using vtols, send out a truck to
build an AA defence structure.
*7 Continue random research while attacking, defending etc ad
infinitum until either victorious or defeated.
Now, as I mentioned above, the WZ AI is simplistic yet gets the job
done, at least to a certain degree. But looking at the above, it is
plain to see that there is no room for improvement, no room for an AI
to grow throughout the course of a game. In other words, the AI is
stuck using it's pre-defined triggers and can't adapt to new
situations, or even recognize new situations.
Pumpkin did good with the tools they had available at the time, but in
this day and age, we can do much better. So, moving on. I'll post the
second part with my comments separately in a second post, so if anyone
reads this, please refrain from posting anything until I'm done
writing the second part, thanks :)
[Edited by ChainsawMonkey]
Ok,
so my thoughts on a real-time adaptive AI.
Looking above, we'll see that everything is done using timed triggers
(at x time we should do this, at xx time we need to do this etc). Ok,
for some stuff this is fine, but throughout the course of a game there
are many different situations that a good AI should be able to take
into consideration, like the layout of an opposing player's base -
where is it most heavily defended? Where are his troops amassed? Is
there a secret tunnel entrance I could use to arrive right in the
middle of his base and cause wholesale carnage at minimal loss to my
own troops? And so on.
Couple this use of triggers with an AI that has no regards for it's
own troops as such and what do you get? Simple, you get a dumb AI. So
how do we get around having dumb AI's to play against? We make them
dynamic, ever-changing, adaptive
Right, so we want a dynamic AI, cool. But how do we achieve this? I
think there are a lot of things we could implement that would give the
AI the fighting chance it needs to actually outwit a human opponent.
So I'll begin with the pre-game loading sequence:
Why have the AI solely based upon timed triggers? Probably because of
limitations on computer hardware at the original time of release. Ok,
so how can we get some more life into our AI? By giving it a
personality. That's right, personality. Obviously not in the sense
that the AI will open communications and start chatting in a pally
fashion about what it did last weekend of course lol. But there is
definitely room for an AI to have personality traits, such as likes
and dislikes. So, we have the pre-game loading sequence. One of the
things that could be generated at this time is the aforementioned
personality traits for each active AI. Let's take favorite color as
being a personality trait (referred to from now on as PT) for an
example.
The game loads, assigning random PT's to each AI (incidentally the
PT's would have some basis in Factions rather than being totally
random). There are 3 AI's and yourself. AI#1 is assigned Blue as fave
color, #2 Red, and #3 Green. Now let's say you decided to choose Red
as your player color. Let's assume that you want to ally yourself with
AI#1, who's fave color is Blue. Now, there would be less likelihood of
this AI allying with you because his PT dictates that he likes Blue,
so you'll quite probably be declined the offer of friendship and
allied status. However, if you approached AI#2, his fave color is Red,
so you'd stand a much better chance of gaining his favor.
Obviously that was an example in it's most primitive form, but I hope
you understand what I'm trying to get across. PT's assigned right from
the beginning of the game like this would instantly make each AI a
unique customer with his own likes and dislikes based upon possibly
dozens of variables including things like Faction (thanks go to Papa
and Flamey for their work on the factions, since the idea of factions
kind of led me to thinking about PT's), environment, even weather
conditions and color preferences. There is a lot of scope here for
making each AI unique.
Now I'll move onto
Triggers, the mainstay of a Warzone AI. Triggers are good, they're
simple and can be checked regularly by the AI with basically no
overhead on the machine running the game. But triggers only go so far,
as will be seen in any game against an AI in Warzone.
Triggers force the AI into doing things regardless of danger or
whether there may be a better alternative.
So let's give the AI the ability to choose it's best course of action.
Say an AI decides to attack another player's base. Rather than just
charging in blindly with masses of troops and units (incidentally in a
straight line that can be picked off one by one), it would send out a
number of scout units (ALS style - ie Air Land and Sea) to the enemy
base and check the layout of the base for weakness.
Now, the enemy base has, say, 3 entrances - one primary front
entrance, a secondary side entrance, and a tertiary rear entrance that
leads directly onto a mass of water. The AI would do a preliminary
scan of defences on each entrance to the base and make a calculated
choice on which entrance would be best to try and crack.
It looks at the front entrance - Not good, Rail and Gauss with Pulse
lasers == heavy losses. So it discounts that entrance for the time
being.
It checks the side entrance - looking better, some HPV and HCannon
with a few Lancers == losses but statistically fewer than the front
entrance.
Now it checks the rear entrance, by using a fast gunboat - minimal
defences, maybe one or 2 Cannons and a couple of destroyers docked at
a port == few losses and an easy attack route through open water.
Now it would depend on the particular forces the AI has available as
to which it would choose in the end. The front is written off straight
away as not being viable for the moment due to heavy casualties being
predicted. So it looks instead at the other 2 options.
The side entrance is viable but what's this? The route to the side
entrance goes right through a narrow valley, meaning possible heavy
losses again due to not being able to mass an attack. Ok, that option
is temporarily written off too.
So it's left with the water route. As luck would have it, the AI has a
number of amphibious assault craft with troop carrying capacity.
Looking at the available information it sees that a water based attack
is the most viable option, using it's amphibious assault craft to
ferry the main attack force right on to the enemy shoreline. It just
needs to send in some destroyers to clear out the enemy vessels ahead
of the main attack force. Decision made, all in the blink of an eye.
However, that 'blink of an eye' would also be enough for you to mount
a defence, if you've got your eyes open and see the scouts sniffing
around your borders
Hopefully you'll see what I'm getting at here, with talks of adaptive
AI. Let's say that in the above example the AI didn't have any Sea
based units. It therefore automatically writes off that option as not
being viable due to lack of appropriate units. So it goes back to the
other 2 routes.
Now technically the second entrance is better since it has less
high-tech devastating weaponry to contend with. However, this route
leads through a valley, meaning a massed attack is out of the
question. The AI might, therefore, decide to risk the losses at the
main entrance, deciding that a massed attack will allow it to destroy
the superior weapons at that entrance faster than it would be able to
at the side entrance.
So you will see, the AI has made an informed decision based simply on
the information it is given by it's scouts And you will also notice
that the general layout of the land had a big hand in the final
decision too.
Again, this option could have also been swayed by the AI's particular
PT's. Say it has a faction specific PT that dictates that it will
prefer to use Tracks over Water propulsion, that could have a big hand
in the decision. Or that the faction specific PT dictates that the AI
is particularly well suited to mountainous warfare, that could have
swayed it in the direction of the Valley assault.
PT's could also affect how an AI lays out it's base. Say the AI has
been assigned a PT that dictates it as being a defensive player - it
would build it's base with primary targets in the best position for
defence and build outwards from there, all the while increasing it's
defensive capabilities.
An offensive player on the other hand would probably start off with
factories near the front of the base for quick assault and worry about
other stuff later.
PT's would also define how willing an AI is to take unnecessary risks.
We all know how dumb the WZ AI is when it sends trucks right to your
front door to build a derrick when you have 6 scourge hard points in
the area lol. A headstrong PT combined with a cunning tactical PT
could provide some interesting scenarios for example, it would be hard
to gauge how it would react because the Headstrong aspect of it's
character could send a rush of tanks unexpectedly right into the teeth
of your defences to catch your attention, while the cunning tactical
aspect could be moving additional troops around the side for a
surprise attack while your busy at the main area of the attack.
As well as determining what PT's an AI starts with, I'd like to see
research topics assigned at the beginning of a skirmish. The AI will
be given a high level topic or possibly a number of topics right from
the outset which they will then work towards getting in an efficient
and organized manner (possibly dependant on PT's again). Thus, the AI
will develop more of a personality as it strives to gain that
particular level of tech, rather than aimlessly researching right
through the available tech tree like Warzone AI does.
And to finish, just a brief note on visibility of terrain and enemies.
It bugs me in Warzone that the AI's all start off pre-informed as to
starting locations. Fair enough, you'll have some vague idea of where
people are, but you shouldn't automatically know who occupies what
location. Therefore I propose that the AI be given a blank map
representing the playing area, with a fog of war that it has to
explore in the same way as a human player (we might not use the fog of
war I know, but I'm referring to the minimap rather than the actual
playing area, essentially meaning the AI starts off as blind as the
human player).
Well, that's it for the moment, I'll add stuff as I think of it. Hope
that gives you something of an idea of how I'd like to see the AI
develop for TWZ
ChainsawMonkey
A brief note on the menu system regarding AI's. There is talk of
having customizable AI, well I think the Personality Traits system
could very well cater for that too. It would be easy enough to set up
a design screen with access to all the various PT's, allowing a mix
and match approach to building a new AI. Also, we could release new
PT's as plug-ins for the game, as well as people being able to write
their own too. The ability to literally mix and match PT's to create a
new AI would be great I think, since it would require no coding on the
part of the user - he or she simply needs to build the AI by placing
the initial PT building blocks Fantastic
Another
brief note:
To avoid areas of high danger.. use a threat assessment map. eg. each
tile given a threat rating based on units/structs near around it.. so
a scouting unit for example will head into the areas (tiles) of least
threat.. and the AI can determine the weakest flank to attack too. :)
spikeles
as
an add on.. we could have a special view mode (sort of like 'z' in WZ)
that shows that in a color scheme.. getting from a green to dark red.
- willis
Yep,
threat assessment maps sound good too, thanks for that Spikeles, and
the 'z' function is good too, but I'd like to see it utilized more
than it was in warzone. As you mentioned Willis, a real-time threat
AOE with red indicating the most dangerous area and green indicating
safe areas. Nice. (ChainSawMonkey)
Ok here's a very simple graphical representation of the threat map in
action. The red circles would represent the optimum firing range for
the weapon emplacements, the yellow circles would be the maximum
firing range and the green circles are safe zones where the weapon
can't fire but a unit is still aware of the weapon's presence.
You'll see that the AI's preferred route takes it straight through the
safe zones, avoiding any risk of being fired upon, but of course this
would depend upon terrain - therefore giving the player more reason to
really think about weapon structure placement in order to catch enemy
units within the cross-fire of a number of emplacements, not to
mention making life more difficult for scout units
[Edited
by ChainsawMonkey]
Ok here's a very simple graphical representation of the threat map in
action. The red circles would represent the optimum firing range
for the weapon emplacements, the yellow circles would be the maximum
firing range and the green circles are safe zones where the
weapon can't fire but a unit is still aware of the weapon's presence.
You'll see that the AI's preferred route takes it straight through the
safe zones, avoiding any risk of being fired upon, but of course this
would depend upon terrain - therefore giving the player more reason to
really think about weapon structure placement in order to catch enemy
units within the cross-fire of a number of emplacements, not to
mention making life more difficult for scout units
[Edited
by ChainsawMonkey]
color-danger
is harder then we think..
brief example, in WZ t1 a machinegun could be considered dangerous...
but how do you compare that to gauss cannon? or a viper unit vs. a
dragon unit. How do we determine what's considered dangerous? i like
your demo monk, n maybe only accounting for range could be a way to do
figures.. but what happens when i spread out a lot (WZ example) of
gauss cannon implants & scourge rocket towers... and on the side
make a huge clump of mg bunkers? it'll think the gauss n scourge route
is safer because its spaced out...and the units are tricked into going
suicide - willis
Just some links:
It's also called "Influence mapping"
This is an entire thread on it:
http://www.gameai.com/influ.thread.html
and an applet for the same...
http://www.ccg.leeds.ac.uk/james/influence/
I'm full of usefull links. :) spikeles
The
scourge towers will have a higher rating.. In the example all the
things are the same "strength". In reality though a machine
gun bunker will have a low value.. say 1.. and a scourge tower will
have 8. So the values around the towers will be HIGHER and more dense
then the bunkers. The bunkers will of course have a rating too. Just a
lower one.
Also. Each unit will have to evaluate it's OWN threat map. Because a
machine gun will do LOTS of damage to a wheeled viper but almost 0 to
a dragon.
Also.. carry the idea a little further.. Imagine the enemy PLOTTING a
course AROUND your air defence to take out your HQ... Just like they
do in real life.. plotting courses around the SAM sites.. it can be
achieved with threat maps! [Edited by Spikeles]