Good AI is Critical for the Next Big Step in Civ Games

[Mark_Everson]

Hi Urban Ranger, thanks for the detailed response.

Originally posted by Urban Ranger
That means the AI has to cover all the possibilities. If something is not in the code/database, it doesn't know how to handle it, and goes haywire in all likelihood. But knowing what to expect in the game is directly related to experience, which isn't something a programmer will have when writing the AI routines. You can tell this by all the expert systems in all the 4X games, they are okay, but ultimately they all rely on cheating - breaking the rules.

What you say is a problem with an AI constructed by very simple heuristics of the if/then type. I agree that that is a standard for AI now, but that is one of the things I think needs to be discarded to get the next step in AI for civ games. We will certainly have a set of heuristics to suggest good starting strategies, but where we differ from the standard approach is that we then passed several strategies to see which fits better in a given situation. What the Clash AI does is it has a cartoon of the world for each type of strategy any uses the cartoon version, which is based on game parameters, to estimate success of a strategy. Each type of strategy, FE "attack area" to try to take control a given area of the map, has its own cartoon of the world that outlines the big issues involved. It's kind of the way people think about it. This is gone over in great detail in our "AI Plan" document in "AI -- the thread" on the Clash forum. There is a link to that thread in the third post of this thread.

Based on the genetic algorithms I have seen Not a whole lot, mind you, but the point is to have the individuals and an environment. If the individuals are the AI routines, the environment will be an arena with a set of criteria where these routines compete.

I am not 100% sure you cannot run the arena in situ, but you will need to spawn hundreds if not thousands of individuals for thousands of generations. This will be much better handled by a meta-game, i.e., run the arena outside the game itself to do the evolution.

Well, it's back to the books for you! That is in fact a typical way GAs are run for engineering optimizations and other things warrior doing your very best to find the absolute global optimal. This is a GAME and I certainly don't attempt to optimize all my strategies in it, why the heck should the AI? The job of GAs in Clash is to just take stabs in the dark looking for "surprise" approaches, that are different from Clash-standard AI, to make the AI less predictable.

A drawback of any evolutionary algorithm is that a solution is "better" only in comparison to other, presently known solutions; such an algorithm actually has no concept of an "optimal solution," or any way to test whether a solution is optimal. This is not entire true of Civ.

When you can play an optimal game of civ I will worry about it. We are only looking for an entertaining and sometimes surprising AI.

[Urban Ranger]

Just read something interesting about Galatic Civilization. Brad said they got 6 AI routines, 5 for the major powers and 1 for the minor powers. Strangely, the 5 are the same thing written by different programmers. They did it to prevent a player from exploiting weaknesses in any one of the routines.

I reckon this can be eliminated if the AI can learn from the human players. In this way, programmers can train the AI so it can take on various approaches (aggresive, defensive, etc.) without extra programming efforts.

A very good example is the old, old Atari game called Castle. Not sure if you can still find it anymore, though.

[mactbone]

Speaking of adaptive ai etc.

Ever hear of EvoChess? I discovered this last year or so and started running it. It's basically a simple chess program that uses "genetic" models to get the best AI. Not only does it pit your generated algorithms against already made "good" algorithms - there is also some migration over the 'net. This was all part of someone's research paper.

I can't find a simple page with everything so here are the various pages (it helps if you know German, I'm sure but most of the documentation is in English)

http://ls11-www.cs.uni-dortmund.de/people/wkantsch/evoChess.htm

http://www.science-at-home.de/index1.htm?/evochess.htm (only in German)

Ahhh, here's probably the best info.

http://www.aspenleaf.com/distributed/distrib-recent.html

Scroll down to Qoopy, about halfway down.

[Panag]

hi ,

after been forgotten for a long time ; http://www-cs-students.stanford.edu/~soren/

its Sorens work on AI , .....


and he is good , ....

have a nice day

[Mark_Everson]

Originally posted by Urban Ranger
Just read something interesting about Galatic Civilization. Brad said they got 6 AI routines, 5 for the major powers and 1 for the minor powers. Strangely, the 5 are the same thing written by different programmers. They did it to prevent a player from exploiting weaknesses in any one of the routines.

I reckon this can be eliminated if the AI can learn from the human players. In this way, programmers can train the AI so it can take on various approaches (aggresive, defensive, etc.) without extra programming efforts.

Hi UR:

I'm no sure I get it. What exact mechanism do you forsee for the AI learning from players? Is it in-game or outside the game (FE by programmers adding new player strategies by hand) that this is done?

Thanks for the references panag and MacTBone. I'll try and check them out at some point.

[LDiCesare]

Soren's AI work is a very clear indication of what he did in Civ3: He says the game must have a small anough search space that the computer can search it efficiently, which means dumbing down game mechanisms so that whatever the ai doesn't know how to exploit effectively should be removed. Someone wrote an article on Civ3 showing that was exactly what happened.

On a different game, MOO3 seems to have gotten out with a pitiful enemy AI (the viceroys seem to be good but not always doing what you expect them to do). The interesting bit of info is that the MOO3 team changed AI settings 2 weeks prior to release because it was too hard, and ended up with a too easy AI. Since that looks like algorithmic rather than text files tweaking, it is quite sad. On another side, GalCiv is supposed to have an efficient AI which can be tweaked by changing just one figure, which decides whether or not it will recognize threats and decide of certain advanced tactics. I'd really like to see this one come out in France soon...