Game development and why most games have "bad" AI

[Chaingun]

Citing the relevant parts from that link:

[11:11] Now, Meier is going to discuss the "Role of AI."

[11:11] He thinks that AI needs to be part of the overall experience but should not be considered a person.

[11:12] He warns that making the AI do surprising things will have bad effects either way. If it does something bad, the player will assume it's stupid. If it does something overly clever, the player will assume it cheated.

[11:13] Meier likes to think of the AI as a metric. It will make the players better and better and will give players feedback. In the case of Civ Rev, the leaders give feedback, since they're the "only friends players have in a game."

As for why games have bad AI, it's actually not just that players want to win, but even more that coding an AI that appears intelligent is incredibly difficult (except for simple games, e.g. chess - and no I don't want to re-discuss the meaning of thw word 'complexity' or discuss chess programming in this thread).

It should also be noted that dumbing down a fundamentally good AI is much easier than making a poor AI better.

I totally agree more difficulty levels = good. Reason being a beginner will not like getting beaten to dust the first time he plays a game, nor does a veteran always win at the hardest difficulty level. Try losing as France on Very Hard as a veteran in EU2/3.

 

[Chaingun]

The problem with making a good AI for a complicated game is not to make it good, it is to make it good AND fast.

I'd say both are problems. Given infinite computing power an AI that fallows the rules of the game will still be finitely good... You'd need to be able to state an optimal solution to the problem of winning the game to exploit infinite computing power. Easy in chess, but try doing it in e.g. Diplomacy ;-).

From practical experience I'll say that making the AI for a non-standard TBS game not being retarded is difficult before you even start optimizing it, though admittedly that's partially due to fact you can't probably shouldn't use algorithms with O(n^2) and worse time complexity. Implementing AIs for checkers and chess were piece of cake in comparison.

 

[Chaingun]

A part of the entertainment in strategy games is having an adversary that isn't brain dead.

 

[Chaingun]

what I don't understand is: computer science has showed many examples of AI-design that weren't considered in game-development. What happened to a 'grown' AI, genetic algorithms and neural networks, why does anybody stays with the old 'IF...,THEN...'-approach?

Genetic algorithms and neural networks have very little application to game AI. Neutral networks are used mainly for pattern recognition, whereas a genetic algorithm essentially is a glorified random search...

Also, define 'old 'IF...,THEN...'-approach' more closely... If you mean finite state machines, yes those are still used. In fact, they will likely continue to be used. It's how you transition between states (e.g. when to go from EAT(object) to ATTACK(enemy) ) that's quite interesting.

 

[podcat]

neural networks are also fairly unpredictable and very hard to train well with back propagation etc when the rewards dont show until very very far in the future. They can be a fun academic toy and I know of one or two projects who use them for very specific problems, but they have no place in most development.

 

[Chaingun]

^

Interesting observations! I agree ideally AI *should* behave so that no game mechanic is rendered completely useless (but this ideal is constrained in practice like earlier discussed in this thread).

Also, like you've noted, it's probably better to give the AI a bonus that is a derivative to time rather than a fixed bonus at start. Though theoretically in Civ more units at startgive exponentially scaling advantages, in practice AI is worse at utilizing them than the human player and thus will have a lower exponential rate, which means any constant factors at start are canceled out given enough time. E.g. (10 * 2 ^ 6t) will be outgrown by (2 ^ 7t).

 

[Chaingun]

I've often thought the same thing, to be sure, but I'm not convinced that's right. If the a.i. in a P'dox game only "samples" the situation once per day and then responds, isn't that essentially similar to a turn-based mechanism? Isn't that what happens when the game pauses for its daily and monthly "rollovers"?

Indeed it is. It's the number of AI frames that vary. (For this discussion, a frame is defined as the smallest unit of time where player input or other events may effect the game world, to which the AI agent must potentially respond. This can be different from graphics frames. ) In a classic turn based game the number of frames is equal to the number of turns (classic as in non-simultaneous TBS). In a real time strategy game player input can occur 60 times per second or more. Paradox games fall in between with a kind of semi-turn system where frames arrive at a user defined rate.

You can "downsample" AI frames, which is what RTS games in general need to do to some extent for many calculations. But then you trade AI quality for execution time.