SE's Best Quality, And How to Improve It

[Arralen]

dmm said:
Anyone else see this article in New Scientist, about how neurotic AIs defeat rational AIs?

Actually, it says the neurotic AI defeats "standard" (totally passive and dumb) AI faster than a plain aggressive AI.

What tells us nothing else than that
a) the standard AI of the game sucks
b) they should have run the test with much more limited in-game ressources and on a bigger time scale: a calculating agressive AI would always beat a neurotical AI in that case. And ressources are limited in one way or the other in 99% of cases. Real life experience shows that ignoring these limitations is recipe for mid- or long-term desaster, not success. The argument of being unpredictable does not apply at all, because the standard AI cannot predict anything at all ... .

[capnq]

I'm reminded of a friend who told me that in the computer version of Risk, when he played against an AI which was set to use the "Random" strategy, it reminded him of playing the boardgame version with me. I used to drive him crazy making attacks that he thought made no sense whatsoever. (They always seemed perfectly sensible to me.)

[narf poit chez BOOM]

...Did you win?

[capnq]

Not that I recall. But most of our games also involved another friend who usually crushed both of us. Which of us got eliminated first largely depended on relative starting positions on the map.

[thebigsilly]

Hi!

Yeah, I wasn't seriously trying to insult anybody. It was a silly goof on my part to capitalize it, but it's also silly that anybody would post just to criticize that; so I made a third silly post just to get passed it.

Anyway --

Well, you know, it's obviously pretty tough to create an AI that can constantly adapt to the player's choices in a way that the player can't exploit. I don't even think that a system like that would be ideal..

The point of my post was to say that, if you had a Space Empires game where the enemy races took predictable but varied routes in terms of military research and tactics, it wouldn't really matter if you could predict and exploit one or even several, because you would always be faced with a radically different opponent somewhere out there. So it would present a lot of challenge without necessarily adding a lot of labor to the process of programming AI's -- you would just have some research techs, ship blueprints and tactical formations flagged at a higher priority for some races than for others. So you would have one race that may never really develop much in the way of shields before it's maxed-out fighter carriers and fighter techs and miniaturized weapons, and this race would go all-out with fighters, so that if you've not got point-defense ships OR super-high-class shields, you're not going to be very successful against them.
At the same time, you would also be facing an enemy with great morale and strength who focus on land combat and ship-boarding, developing those techs and building those kinds of units/ships before they worry about space-based weaponry.
And also another enemy who builds-up defense research as quickly as possible, cramming as many mines, satellites and stations around their warp-points and colonies as they can manage.
And so on, and so on. And being faced with all of these radically different opponents at the same time would compel the player to really try to only take on the enemies that he's prepared to fight against. Even if you can exploit the particular tendencies of a handful of opponents, it really wouldn't be possible to develop yourself in a way that could account for ALL of your opponents' traits. Just breaking down the general AI's building/research schemes into race-specific schemes would add new gameplay dimensions to almost every aspect of the game - for one, diplomacy, which is kind of a weak spot in single player SE despite there being tons of options. Now it would really matter that you're able to placate specific races - what if their tech is particularly well-suited to besting your own, because you've been developing to confront a more immediate threat?

I don't really think randomness needs to enter into it for the kinds of things that I'm talking about to work. Only variety.

[dmm]

I haven't played with SEIV's or SEV's AI scripting. Are there triggers? I've played with other games' AI scripts, and it seems like reacting to events is always a huge missing element differentiating AIs from humans.

For example:
In SE, if a human suddenly encounters a race with good missiles or fighters, he will either
a) make PD a research priority, or
b) desperately trade for PD, or
c) go into defensive mode, or
d) assiduously avoid that race, or
e) make extra-nice with that race, or
f) go all-out to destroy that race quickly.

Another example:
If a human is losing colony after colony to one player and the situation is hopeless, he will either
a) quit , or
b) surrender, or
c) try to negotiate a protectorate deal, or
d) get vindictive by going kamikaze, or
e) get vindictive by surrendering to another player.

But the AIs just seem to go on with whatever strategy that was pre-mapped out for them, regardless of the circumstances. Triggered behaviors would solve a lot of that problem, and they could be tuned for each race to match that race's personality.

[Captain Kwok]

The problem with creating a reactionary AI is how do you determine how they respond in an environment with multiple empires and get them to direct the appropriate response versus a specific empire. Human players can manage this a lot better and that's a big advantage they have - matching up counter strategies versus each opponent.

[dmm]

what do you do? Here's an idea, feel free to bash it as much as you like. Suppose the following:
a) AI players are plentiful: 25, maybe even 50
b) mediocre planets are plentiful
c) asteroid fields are plentiful and worth developing
d1) AIs don't like colonizing in "their" systems
d2) but they don't totally freak out about it
d3) unless it gets "too much", then they freak out
e) cheap small exploratory probes are default tech
e1) maybe cloaked? extra-fast? solar-powered? unmanned?
e2) definitely unarmed
f) AIs form lots of gradually-improving alliances
g) AIs react negatively toward attacks on allies
g1) reaction is mild if ally is casual acquaintance
g2) reaction is strong if ally is firm friend

If this could be implemented, I think that (if the parameters were set correctly) it would set up a galaxy that would gradually become well-populated, technologically advanced, very interconnected, and incredibly unstable. Woohoo!!

So, what do you think? Is this do-able right now? Would it work? Has it been tried already?

[capnq]

dmm said: f) AIs form lots of gradually-improving alliances

g1) reaction is mild if ally is casual acquaintance
g2) reaction is strong if ally is firm friend

These threee look like the hardest to implement. How do you define "improving", "casual", and "firm", in terms the AI understands?

[Phoenix-D]

Civ 4 does it decently. They make a list of things that the AI likes (tributes, you beating up on his enemies, time allied) and things it dislikes (being nuked, bad demands, being allied with his enemies) and weight the AI's attitude and alliances by that.