Interview: SOREN JOHNSON ON AI

[MarkG]

hope you find it interesting

http://apolyton.net/misc/interviews/...njohnson.shtml

[yin26]

Thanks Markos and Soren. I think a good number of us here feel that a great AI is the beginning, middle and end of a great Civ3. I look forward to seeing the learning algorithm in action.

[Imran Siddiqui]

Need a news item written?

[Immortal Wombat]

Cool

Sounds promising, though not inspirationally awe-inspiring. It should make for decent variety of play at least. Not much there about how well the decision-making improves the AI in terms of how hard it will be to beat.

Yeah, promising, but not amazing. Lets hope the AI has got more up its sleeve that we won't see until its "learnt" a few ways of killing us.

[Lawrence of Arabia]

Looks like whoever was saying that the AI would suck just got shafted. And turning off those pre set features is the first thing im doing.

[yin26]

Looks like whoever was saying that the AI would suck just got shafted.

Actually, his interview wasn't detailed enough to make any conclusions yet. I've seen plenty of 'learning AI' do absolutely nothing but suck in previous games (Submarine Titans being one).

[Immortal Wombat]

Originally posted by Lawrence of Arabia
Looks like whoever was saying that the AI would suck just got shafted.

Not really.
So your governors can choose not to build archers if your at peace, and miles from anywhere. Whopee. Can other civs handle all the resources they need, how to get them, why to get them, when to get them with a longer-term strategy than "oh damn, the human just declared war on me with no warnin, maybe I should get some coal"?

I hope like buggery I'm wrong, bu a good all-round AI is damn nearly impossible to cram into a few seconds of CPU time. (or even a few minutes of it)

[dennis580]

Well at least we now know the Civ III AI can cordinate ground, sea, and air units effectively in a invasion. I think this is a very big step and gives great promise for the rest of the AI in the game

[SerapisIV]

The AI using carrier-based bombers to support marine invasions?

Let me recheck this, the Civ3 AI is actually capable of using carriers for invasions, then actually using marine's for sea invasions, not glorified infantry?

God I hope all of this still pans out when the game shows up on my hard drive in the next week or so, the promises of a much better Civ AI is extremely enticing.

[The diplomat]

Some of you just want to be pessimistic, don't you?

We obviously won't completely know how good the AI is until we actually play the game, but this interview is very promising. If the AI can actually do what Soren claims, then the AI will be very decent to say the least!

[Skanky Burns]

Well, the good new is that if the City Governor AI is smarter for controlling your cities, its also going to be smarter in controlling its own cities. Your governors take into account your overall policy, and learn from your input. This is also good

[dennis580]

Yea this will also be good from the computer AI. Obviously thier's going to be a limit to how much communication the computer leader and the computer controlled city's have. The cmputer controlled cities need to have good idea of thier leaders overall plans.

[The Rook]

Has there ever been a really great AI ever? In any game? Computers are just not that capable of making decisions on such a large scale. What computers ARE good at doing is repitition. That's why they can be taught to play chess well. Chess has a limited number of moves, so the computer "brain" just goes through them all until it finds the best option. I think the same thing could be done with Civ, but on a lesser scale. During the human turn, the computer Civs can be planning, thinking ahead, and formulating a strategy, then on its turn, it acts on the strategy it came up with. Each turn it can adapt and improve on the strategy.

I just think during the long human turns, it could be doing number crunching and whatever.

[CobraA1]

The AI sounds very promising.

• Multiple-level thinking
• Personality
• Some customizeability, including personality and priorities of AI and the ability to start with a blank slate for those who dislike the personalities.
• No pre-scripted patterns
• Learning algorithms
• Emergent Behavior

Even if it turns out to be stupid, it looks like its a huge leap from Civ II, and also looks like it's going to be fun to play against.

Emergent behavior, however, has me a little scared . . .

[squid]

Originally posted by yin26


Actually, his interview wasn't detailed enough to make any conclusions yet. I've seen plenty of 'learning AI' do absolutely nothing but suck in previous games (Submarine Titans being one).

From what I've read (and I took some AI in college), he is stressing an adaptive AI but not neccesarily learning.

learning would imply it learned from the consequences of action and from your actions.

it sounds more like it is using static evaluators to access positions than by doing any sort of learning activity. For instance, "The romans are too powerful. We must attack Rome if we are strong enough. We are strong enough. What is the closest strongest city? Thebes. What is their closest weak city? Remus. Let us attack from Thebes. Thebes, sent forth some units. Units, approach the city of Remus. Phalanx: yes, I will do so. What is the closest path to Remus? I will follow it. There is a unit in my way. It is threatening. I will not survive. Must kill unit". I am trivializing things, but yeah, his hierachial system sounds much like this.

As for expansion versus conquest versus introspection, the AI could easily decide which was a deficiency, and then use the "leader module" to order down the tree (to cities, governers, units, etc) appropraitely. Not that hard.

Anyhow, of course it's not hard. The question is it done well. If I can see bombers softening a landing area before landing a transport, I will be impressed.

It seems Soren has some good ideas though. I was assuming they would have had to consider repurcussions several turns in advance. That does not seem to be the case!

[cassembler]

Well, I remember playing SMAC after years of Civ2, and it was pleasantly surprising- for a while. If the AI can do much of my micromanagement, and it looks like it can, then I can create scenarios that promote the AI behavior I'm looking for...

I don't think the AI for Civ3 will be in the catgory of "bad." I think it will be at least "fair to midland."

[StratAll]

Quote :
Hence, the city AI is not based on pre-scripted patterns, such as Spearman, Temple, Spearman, Granary, Archer, etc. Instead, each available build item is given a value depending on the AI’s environment, and the city simply builds the item with the highest value.
End quote

I have no experience with making mods, I only played a few in CivII but I guess this is very good news for mod makers. The AI will now fully understand the potential of new units created and will be able to weigh their value against one another based on its own situation.

[Lorizael]

I think the three best things mentioned here are the facts that the AI coordinates attacks

Works on different levels from all the micromanagement to the big plans of exploration and domination

And does not work on pre-scripted patterns.

While not very detailed the information here definetely gives me hope for this game. Those are the areas that I thought needed to be improved to make the AI good.

And of course my opinion is held in such high regard by the people at Firaxis...

[Deathray]

Originally posted by The Rook
Has there ever been a really great AI ever? In any game? Computers are just not that capable of making decisions on such a large scale. What computers ARE good at doing is repitition. That's why they can be taught to play chess well. Chess has a limited number of moves, so the computer "brain" just goes through them all until it finds the best option. I think the same thing could be done with Civ, but on a lesser scale. During the human turn, the computer Civs can be planning, thinking ahead, and formulating a strategy, then on its turn, it acts on the strategy it came up with. Each turn it can adapt and improve on the strategy.

I just think during the long human turns, it could be doing number crunching and whatever.

The Simtex games (Master of Orion & Master of Magic) had pretty damn good AI. Both beatable though, even without use of cheese.

[Solver]

Great interview! Go AI and go Sid! 3 days more till Civ 3.

[Locutus]

From what I read, there are many good things about the AI in Civ3 and some mayor improvements over Civ2, but what worries me most is the lack of customizability. Sure, you can change their personality and (some) priorities but what if it turns out that the AI always neglects culture or if it builds lots of units but uses them in completely the wrong way or can't deal with flaws in the game-engine (ICS anyone?)? I'm not saying any of this is actually the case but I'm sure we'll find flaws in the AI soon enough. From what I read, there is no way for modmakers to fix these flaws or to adapt the AI to deal with specific situations (read: scenarios). In CtP, most AI routines can be accessed and changed through text files and a scripting language and thus be totally transformed, looks like something like this won't be possible in Civ3. Since there's no multiplayer this could severely damage the replayability value of the game in the long run...

[old dog]

"These personalities are fully customizable through the editor, so creating AI’s with different priorities is quite easy. " I am very relieved to hear this. This will have a large effect on replay value. No great surprise, though.

"Users who dislike preset personalities can turn them off altogether and let the AI adapt to its environment from a blank slate. " This is EXCELLENT news. The twentieth time I play against Shaka, it will get kind of old knowing that he will be bloodthirsty. Predictability is the enemy of replayability. Furthermore, it may well be that these "blank slate" leaders will be the most effective competition. Who can forget the Mongols in Civ 1, alone on some island, spewing out low level military units, boiling about looking for someone to fight?

"The number of civilizations in a game definitely affects AI behavior... The AI adapts to its environment, and the number of civilizations in the world is simply one of the many factors used for decision-making. For example, if a game was started on a Huge map with only 6 civs, the AI might emphasize expansion to take advantage of the extra space available." If completely true, this is the best news of all. Personally, I am very attracted to the idea of a crowded world, where 16 civs start on an average size map. I like expansion to be something you have to work at... However, for this to work as a competitive game, the AI needs to adapt to the situation... And, of course, for scenerio writing, this aspect of the AI is key. How many times have I started playing someone's interesting looking scenerio but become bored because the AI was brain dead to the situation and my strategy?

Of course, there are lot of important things that Johnson did not touch on:
1) How effective is the AI at attempting each of the types of victory? Can it assess its chances of success at each, and then pursue that strategy?
2) How well will the AI play the resource game? It's my sense that this will be the true centerpiece of Civ3, and the ability to obtain resources for yourself while denying them to your most important rivals will be the overriding theme of gameplay. If the AI does this astutely, it will go a long way towards making up for any (likely) deficiencies in waging offensive war.
3) Johnson says that the AI reacts to the situation, but does it react to its own success and failures? If two ships of soldiers sent to attack a particular city fail to make a dent, what are the chances that the same number of units will be sent to attack the same city again? If the ships were sunk, will the AI try a different route or a stronger navy?

[Comrade Tribune]

Originally posted by MarkG
hope you find it interesting

The understatement of the year.

[Solver]

Don't know, but thow shole interview made me that much more optimistic about the game.

[Comrade Tribune]

Originally posted by The Rook
I just think during the long human turns, it could be doing number crunching and whatever.

Good point. I sure hope the AI will utilize the time the player takes to make his decisions. Afaik, it didn´t do so in any previous civgame.

[Thue]

Furthermore, the different AI levels communicate between each other. For example, the unit AI controlling a tank might ask the leader AI which enemy city would be best to attack. After receiving a response, the unit AI would determine its own best path to the destination. Thus, the AI is able to organize a large group of units while keeping the discrete game decisions at a lower level.

If the AI is worth anything it decides the best path by making a move cost map before making the decision, to factor in the precise distance.

[Grim Legacy]

Originally posted by Thue


If the AI is worth anything it decides the best path by making a move cost map before making the decision, to factor in the precise distance.

I think that calculation is made on the higher level that sets the goal. The "unit AI" takes care of more micro-esque decisions.

It's unlikely that distance is never taken in to account in a game like civ. There are strong indications that the programmers used distance as a factor in Civ2 at least.

[HalfLotus]

This sounds promising. I like that the civ personalities can be turned off. People seem to forget that we've been playing Civ2 for five years now, and have learned the AI inside and out. Seeing as how the AI has been completely redone, we'll be starting from scratch again. Surely we'll have a few months of pleasants surprises from the AI.

[Rhysie]

Originally posted by Comrade Tribune


Good point. I sure hope the AI will utilize the time the player takes to make his decisions. Afaik, it didn´t do so in any previous civgame.

I hope that if it does, this only happens at higher difficulty levels - Otherwise a warlord game played by a very slow, new player who takes their time and looks at everything before even moving a unit might find the AI they face harder than someone who is very quick and judges the situation without too much delay, making snap decisions which allow them to have very short turns could be facing a weaker AI on a higher difficulty level.

I understand that many players want a challenging AI, but I, along with some others, would like an AI which I can play against in a relaxed manner, not micromanaging everything and occasionally doing things I want to rather than because they will improve my winning potential, and still be assured of an eventual win.

[faded glory]

great! but im still pessemistic. Ill wait till I play monday