I´m not wishing to do the system because I´m not actualy making a FPS or anything that may require it now.

But I just wanted to leave my thoughts about my project of True AI for a game.

//////
But before explaining my theory, I would like to know if there are actually some games out there with Learning AI on their NPCs. I would like to try them out BTW.
//////

Ok, my idea is pretty simple, it´s basically a trial and error with tactics ranking and some linear thinking. Let the NPC use a weapon and try different approaches with it. The NPC automatically rates the tactic with an integer value, depending on how much damage it actually did to an enemy, and if it got damaged itself. Also, he should watch out for allies. Then it should take some time to classify tactics(burst, Single, Area Effects).

Of course, he should add some conditions to using different tactics depending on its enemy condition. This might be something nice like for example, the NPC actually does not know if his enemy is immune to some type of damage, but actually trying it he might learn the consequences.

Basically, you let it play two hours and it might be the smartest bot ever or maybe not

I´m not going to do it for any game by now, but I though it would be pretty cool for a FPS game.

In my mind there are a few different things that go into making AI believable:

• It has to have some inherent smarts

• It has to retain information

• It has to extrapolate that information

• It has to have some "play" to it -- experiment every once in a while

• It has to, occasionally, be stupid -- not very often, but it needs to make some mistakes

The first three are purely mechanical, the last two are a matter of tuning. I think these are easier to do in an RTS, but I think you can accomplish the same things in an FPS.

I think the things you need to figure out are:

• Who is gathering the information

• Who is receiving the information

• What they already know about the enemy and environment

• What information they're gathering

• What kind of lifetime the information has

• What actions depend on what kind of gathered information

For instance, you could have a group-based AI where every individual agent reports what it finds (direct encounters, ammo/health pickups, spent rounds, sounds, footprints...) to the group leader, who reacts by issuing orders and stores the information in an influence map for pathfinding later. That information has an exponential decay so that old information becomes irrelevant and eventually drops off.

Thats like saying "Don't expect a human to learn how to play a FPS a few hours after its born". Takes our neural net years to develop the capacity to do what it does. I can only imagine it'd take a simulated neural net as long or longer (using current technology at least).

Computers and the programs don't start out with incomplete core's either. If the computer the AI is running on and the AI itself had to grow and evolve like the human brain did, I think even in those special cases it'd take time

Indeed. Theres no way you'll get a "learning computer" to learn what it needs in so short a time (imo). least not with current technology.

http://nerogame.org/

What do you think those guys I talked about did?

That NERO game looks like fun! But instead of playing the game, I'd download the learning engine (If I were you )
http://nn.cs.utexas.edu/project-view.php?ProjID=68#software_list

[quote]...[/quote]

Usually with considerably smart AI, people don't understand what the AI is doing, and don't attribute actions to intelligence. If you have gaming applications for something you are looking into making, keep the player's perspective in mind when making it. Make the formations distinct, have them shout to each other, etc. Make sure the player can get what's going on. Your system sounds like a good system for breaking up monotony to a one-versus-many FPS if done right.

I thought they discontinued the Linux port k3b from kde3 is much better anyhow (last time I tried the kde4 version I couldn't get it to actually start burning a disk).

One way to make a learning FPS AI is to keep track of the locations the AI was Killed, or damaged and the locations the AI gave damage or got a kill. By gravitating towards the places where they got kills and avoiding places where they got killed, the agent will at least develop a sense of where to stand in a particular level. Also you can plug in data from human players to allow the AI to learn from other people's actions.

I know Steam is keeping track of this data for Team Fortress 2. Looking at their death density maps is pretty interesting.