Quote: Original post by MrTwitchy
Our job as game AI programmers is to help create the most entertaining product, not to make the best AI.

And yet these are not mutually exculsive concepts, are they? And yet the industry keeps pitting them on opposite sides of the argument. Good AI can be entertaining. Ponder that for a bit.

Quote: Original post by InnocuousFox

Quote: Original post by MrTwitchy
Our job as game AI programmers is to help create the most entertaining product, not to make the best AI.

And yet these are not mutually exculsive concepts, are they?

They're not mutually exclusive concepts, no, but my point was that some games deliberately choose to use hobbled AI because that's one of the ways they're creating fun for the player. This may not appeal to your tastes, but it's just as valid a method.

It all comes down to entertainment in the end, as we are creating a product designed to entertain. Some games are designed such that the AI is most entertaining by being exceptionally realistic, highly tactical, etc, etc. In some games the AI is most entertaining by displaying designed in weaknesses and a certain predictability.

I guess it depends on your definition on good/bad/perfect to easy is bad, to hard is bad. Now the OP has created a bad AI because there is an obvious and easily exploitable hole. So it may be fun in the short term but I bet it gets old quickly.

Quote: Original post by ToohrVyk
...it would be enough to have three low-level monsters with perfect AI to kill the player.

Perfect AI is not good AI, either. That is another ridiculous misconception. One of the caveats to von Neumann's Game Theory was that it assumed that the decision maker was entirely rational. That is where it came up short as a modeling tool. Humans (et al) are not completely rational.

When we design computer AI with perfect knowledge, perfect calculation, and perfect execution, we are creating an entity that is not "human-like" precisely because of the fact that we are creating something perfect.

The point is, for too long we have been focused on "solving" a problem with AI. To solve the problem from a scientific, mathematical or engineering standpoint means to maximize it. (i.e. perfect it) However, by doing so, this became no longer fun since any given bot could head-shot us in an instant, etc. So, in a way, we abandoned that quest because we were told "perfect != fun".

What if we were to construct our agents in such a way that we weren't striving for the very non-human, non-fun perfection and, instead, strove to model something that was realistic in the sense of being sub-optimal and even fallible. Not in the contrived sense - that is with "exploits". But in a reasonable fashion.

For example, at a GDC many years ago, I shared a beer with someone who had worked on the AI for a particular NFL game I had played. I told him that my observation was that you could march down the field entirely by doing 10-yard hook patterns. The DB would always be just off the ball as the WR cut back. He looked baffled when I suggested that there could be a fix to this.

I went on to explain that (in real football), after a number of times, the DB would tend to jump the route and try to cut off the hook route. He argued that the human player could "exploit" this feature by then running a fake hook and go right past the DB after drawing him in. It was my turn to be baffled. It was obvious that he had not watched a lot of football.

The point was, the issue that I brought up was an "exploit". The player could take advantage of something that the designer/programmer had obviously not put into the DB logic. The caveat that he thought was an "exploit" is a truly legitimate play based on bluffing. You can't watch an NFL game without hearing such things like "run to set up the pass", "drawing the defense in", "taking the underneath route that they are giving them", and, more relevantly "setting the defender up by running the same route a few times."

In both cases, the DB made an error. In the first, he made the error of doing the same thing every time and not adjusting to the exploit that the player had found. In the second error, the DB got drawn in to doing a legitimate reaction to a legitimate strategy just like DBs do every single week in the NFL.

For the player, the difference is best illustrated by the following sentences:

* I beat the game by finding something stupid in the AI that the designers didn't account for.
* I beat the game by finding a strategy that took advantage of the AI's realistic response.

The first is solving a puzzle. The second is playing football. If you are creating a puzzle game, have at it. If you are trying to create a football simulation, then "bad AI" is disappointing.

By the way, after all this I may as well pimp my book, "Behavioral Mathematics for Game AI". Hopefully it will be on the shelves in time for GDC. It addresses exactly the above issue... step out of the hyper-focus on algorithms and write behaviors instead.

One time I pondered if being a general came down to rock-paper-sissors. That thought deflated my interest. There was this really cool game called "the Perfect General" where you had to predict, a turn ahead, where the enemy was going to be, so you could place artillery barrages. My game bears a passing resemblance to the Perfect General and when I decided I didn't have anything better or different to offer, I just left my game as unfinished, to be used as a tutorial for those who might want to use it to make their own. It shows-off what LoseThos has to offer.

Relevant reading

Quote: Original post by InnocuousFox
Relevant reading

Thanks for the link.

My partner, who's an old-time wargamer, finds our game much more enjoyable every time the AI improves. It's good enough now to make you watch your flanks, pay attention to the center, and be careful of powerful attackers "sneaking" up on you. (Much of that improvement is due to reading on this forum.)

I enjoy good AI... that I can beat ;-) There was this stupid chess game which was grand master level or something. That was no fun.

To quote Soren Johnson (Civ 3 & 4), "if the player doesn't win he should at least understand why and learn from it."

Quote: Original post by losethos
I just did a tank game on a hex map. The AI is brain dead--all enemy units move toward the nearest of your units and shoot at the nearest. I like exploiting knowledge of the AI to kick it's ass by placing a bait unit to direct its movement.

Is it just me or is bad AI more fun, sometimes? Who doesn't like winning by a big margin. Many times I've played against humans and decided it wasn't as much fun ;-)

If you're in the mood for doing some Ai for fun, you can improve the AI on this: LoseThos Tank Game

Bad AI might be funny at first, but turn to boring (and worse) fairly quick.
Winning by a big margin ?? What about the second, third, fourth time you play it (if it isnt in the trash already by then). Humans too challenging ?? Well good game AI can always be made dumber, and you should be able to dial it down
when you feel like an easy game. But then when you feel like a challenge then you can dial it up.