A.I. thoughts

Post by **Balthagor** » Feb 23 2004

I got talking a bit about the AI in the â€œopen ended playâ€ thread and thought I could get some opinions on how aggressive the AI should be in general.

As I mentioned in the other post, we are building the AIs working from the premise of agendas. Some AI players will be out for Conquest, others for Reunification. The agenda of the AI player can change based on in game events so that if you really start to beat on an AI with an Intimidation agenda, it could shift to Conquest since youâ€™ve ticked them off. I also think that Conquest should be the most common agenda. The story line has created a world where everyone has become desensitized to all the violence and believes that the only solution will be one Supreme Ruler to rule them all. (â€¦was that a veiled movie reference?) A world like this would cause many people to think â€œhey, why not me?â€ and we also have many new â€œnation leadersâ€ whoâ€™ve just gotten control of a new country. Power is addictive and a little is rarely enough.

Now, Iâ€™ve played enough games with single minded AIs that do nothing but attack the player ignoring each other to say that is certainly not what I want to see here. The more players in a map, the more variety of agendas there should be so that the player gets a dynamic like what weâ€™ve seen in multiplayer games. Now this does mean that the AI will need to be able to backstab and make opportunist moves. One of the recent test games I played, the two beta testers decided to try and rush me off the map. The three players in Michigan are very closely balanced so I was seeing about double my entire force coming at me from two directions. I got chatting telling them that they where successfully crushing me (and they where) but this caused the Michigan player to decide I was no longer the threat and turn to start attacking the Ohio player. Obviously, you canâ€™t chat with the AI, but the diplomacy system will allow you to send â€œconcessionaryâ€ offers that will give the AI a feel for your intentions. I hope to see an AI that can understand when itâ€™s on the ropes, when itâ€™s in a good position to strike and when it needs help from others.

Anyway, some of this was a little rambling, but hopefully gives an idea of what weâ€™re working for and will get some feedback.

FastBoy101 · Post by **FastBoy101** » Feb 23 2004

Thanks for the information, it's nice to know where it's going once in a while

Geta · Post by **Geta** » Feb 23 2004

I write computer game AI for a living, although at this time I am not working on a strategy game like SR2010. In the past when I did work on turn and RT strategy games, I thought that if given the time, I would have put more effort into "utility evaluation" of different courses of action than I had previously done. Since you describe an "agenda driven" AI approach, that suggests that the goal or purpose of the AI has already been determined, and then it is now up to the AI code to evaluate the various courses of action to take, that get the most utility (or are the best) to take to achieve the agenda.

What I am suggesting is more high level as opposed to "nuts and bolts" of how to do it. So I am suggesting that developing a selection (or menu) of actions that could be taken to accomplish an agenda, from a specific point in time, and then perform significant evaluation of their utility. Once rated, additional criteria, even determined by the agenda, could be applied to determine if the resulting utility values are considered "good" or "bad".

The advantage to this approach, should be an AI that develops courses of action that are not only sensative to the information it has on its opponents, but also would vary from time to time.

Just my 2 cents worth.

JXai · Post by **JXai** » Feb 23 2004

I'd like to see an AI that makes some crucial mistakes, such as selling out to overextend its resources on offense only to weaken themselves at home -- or one that tries so hard to be at peace, it leaves itself open to attack.

As far as aggression goes, I think the AI can be as aggressive as you can make it, but NOT if it does so because it has an inhuman knack for empire building. Again, I'd like to see an AI personality attack a little earlier than it "should" and attempt to catch the human opponent off-guard...but OOPS...the human defeats the attack and the AI is now up $%&^ creek.

Those characteristics would be interesting to me.

Post by **BattleGoat** » Feb 24 2004

There will be different AI personalities involved in the game, so yes you will see the Aggressive guys gambling a little more militarily then would be prudent, and then again the Defensive AI's might miss opportunities that they should have taken advantage of!

djtrix · Post by **djtrix** » Feb 24 2004

I enjoyed the option in "Galactic Civilizations" where you could set the evil vs good level and the intelligence level of the AI for open-ended games. Although this would obviously not work for campaign games, I'd love to see it in sandbox mode! Sometimes it's nice to set up a very favourable game and conquer the world with an iron fist. Other times it's nice to set up a difficult game where you struggle for victory. Having the option is definately a plus in my books.

Ashbery76 · Post by **Ashbery76** » Feb 25 2004

I don't like the idea of aggresive and defensive A.I personalitys in military matters, surely this should just be based of the advantage of the army size.The A.I that has a very small army compared to the enemy should set up defensive positions and try the inflict attrition.

Another thing i hate to see is an A.I declare war when it has an army that is terrible! and has no chance to do anything.I only want to see A.I go on the offensive and DOW when it has forces with similar numbers to the enemy..

Post by **Balthagor** » Feb 25 2004

I'll assume that smaller numbers but better equipped still matches your criteria. I see what you mean and I do agree that the AI should not just declare war for stupid reasons, however I do want to see the AI be somewhat unpredictable. If you and I where playing a multiplayer game and you consistently refused to sell me technologies and units and blocked all my treaty offers, I might just attack you for no good reason. People do that. I feel that the AI needs to react to events, but its reaction would be based on its agenda. You attack the ally of a â€œConquestâ€ agenda, if youâ€™re stronger, I think there should be a chance that this AI turns on its ally and sends you diplomatic offers. Attack the ally of an AI with an agenda for reunification, and it should stand behind its original ally to the last unit. If done properly, the AI will also added what it knows of the opponent to the equation when choosing an action, all while calculating what best suits its agenda.

Juergen · Post by **Juergen** » Feb 26 2004

The AI shouldnt be too tame,the player should be affraid of it at any time.
This way there is a good reason in SP to build a protective army instead of "rushing" your economy and technology.
This way there actually is something to balance.

If that means that the AI attacks even when the odds are not that good,this would be far better with me than an AI which does nothing.
And who knows?
Maybe the AI wins,even when the war didnt seem so good from the start.

Granted the AI should also deal with its own land and improve its own economy.

What I would also like to see would be an AI who actively uses WMD.
It should be capable of putting them into good use and also know when to use them.
It should use actual "counterforce" and "countervalue" strategy.

And also in this case I would rather see it drop some bombs more than less,the player should have a good reason to fear the AI WMD arsenal.
Knowing that the AI might just sit on these things or that it wouldnt use them first wont create this fear.

Geta · Post by **Geta** » Feb 26 2004

I would like to see the SR2010 AI opponent stage actual attacks in force, and not just dribble units into the fray. Save up, stage units, attack from multiple directions with a valid combined arms doctrine. Then occasionally send in an all armor recon in force, or an all helicopter raid. Easier requested than done.

Post by **Balthagor** » Feb 27 2004

Yes, a cheating AI is certainly something we all find to be a poor cover for a weak AI.

Post by **George Geczy** » Feb 27 2004

Some interesting comments in this thread.

Certainly, as mentioned, we will not be allowing a 'cheating' or all-knowing AI. It has all the same awareness and knowledge as a human player would.

I haven't made the final decision on exactly how to 'tone down' the AI to make it a simpler opponent when the game setting is 'easy', but I don't like the idea of having it make stupid mistakes. Instead, I'm leaning towards a 'pacifist' AI for the easier settings, as well as the old standby of giving it (or the human player) more or less resources and money to start with.

As much as there is a concern about making the AI really good, there is also the fear that the AI could badly maul a first-time player and turn them off from the game ("don't lead with your artillery! watch your supply lines! Ouch!"), so we'll need to be able to have ways of addressing that.

The other balancing element could be the other regions in a scenario - in an 'easy' game, they could be set to lean towards supporting the human player; in a 'hard' game, they would be more likely to support each other and have a dislike for the human.

-- George.

Juergen · Post by **Juergen** » Feb 27 2004

Interesting points.

1)I agree with you George,its hard enough to make an AI perform as it should.
Trying to make an AI that makes "mistakes" that are convincing and not just plain stupid might ultimately distract from making a good AI and thus do more harm than good.

2)I havent programmed an AI myself but I take it that its generally harder to make multiple difficulty settings for an AI than a sinlge one,right?
And after making a tough AI it would be hard to derive more easier setting for the AI?
Is there a decision between an AI thats too easy and an AI thats to hard?

3)I would consider the old "AI players get together to beat human" trick not much better than a cheat.
Its quite an old hat (since CIV1),I have seen its effect often enough.
Granted it does make the game harder but it does so by changing the numbers against the player and NOT by playing any better.
A resource cheat does the same but in another way.

Altough,I would be fine with a more "suspecious" AI.
An AI of that would react far more hostile to your ACTIONS.
An AI that reacts different at what you (or other AI players) do rather than what you are.
Where the AI on the easy level would do nothing to question or counteract your actions the AI on the hard level would try to figure out wether you are fooling it or not and when in doubt it would assume the later.
And then act accordingly.

<font size=-1>[ This Message was edited by: Juergen on 2004-02-27 05:45 ]</font>

<font size=-1>[ This Message was edited by: Juergen on 2004-02-27 05:50 ]</font>

Ashbery76 · Post by **Ashbery76** » Feb 27 2004

On 2004-02-26 19:16, Hanibal wrote:
yes i do agree AI should have personalities from "offensive" AI to a more conservative one

So you think an A.I with one infantry division should attack a nation with fifty due to a personality!.Suicide A.I is never good for a game and not based on reality, saddam would never have attacked kuwait again and didn't, even though he had a hostile personality toward them.