AI and Why?

We had an interesting discussion about AI in the irc channel earlier this week and I wanted to relate some of it here in hopes of sparking a larger discussion.

Now, for the record, I have no experience with AI programming. I also know but a little of theory. So with that understanding, I am not trying to offer solutions but rather observations.

AI is getting smarter everyday. As computers get faster, games have more and more resources to throw at making the AI better. From what little experience I do have, it is generally not hard to cook up a way to make your AI smarter. Instead it is hard to cook up a way to make him smarter given limited computational resources. The good news here is that this problem will solve itself given time as computers continue to get faster. The graphics in a gmae are rapidly moving off the processor to gpus and in a few years we will have all the processing power needed to make our physics as realistic as we'd want. At that point all extra processing power can be plied towards the AI and AI programming will truly come into focus.

Now while this is nice and good and all, I already find that the AI in modern games have very little difficulty beating the crap out of me so I am not overly concerned with making them smarter. As I've heard it said, you don't need to make the AI smart, you just need to make them smart enough to give me a good fight and then to lose (I think Phil Carlisle said this).

Rather than worrying about making the AI smarter, I find myself concerned with making me and my audience care about the AI. After years of playing against humans and bots, I generally find that I get no enjoyment out of beating a bot. There is simply something more infinitely satisfying about playing with and against humans.

I think there are two critical elements needed for the player to start to care about an AI character. Usefullness and communication. We've all had to babysit our AI teammates from time to time and it continually reminds us that we are playing with a computer controlled automaton. A lot of this is the "smarts" I referred to earlier and I proposed, this has and will continue to get better.

There is a third element needed for the player to emotionally interact (love or hate) with a bot: personality. However, as Joe Maru pointed out to me, personality is simply an expected set of responses that the player gets to know, whether it be that the bot likes to make jokes about your mother or whether he tends to hang back in fights. As Joe pointed out with his experiences from Think Tanks, the "personalities" of their bots basically fell out of their pre-programmed behavoirs. People know that this bot is a sneaky sniper and that other bot is a suicidal nutcase. I think that it would be fairly easy for a game like Think Tanks to play on these personalities by given them voices and extending their revenge system into something more complex.

As a brief aside I would suggest that a revenge/grudge system is not quite enough to give the AI human-like behavior. Often times in multiplayer games you will find yourself forging temporary alliances with other players. Usually, these are unspoken. Sometimes you will just find yourself pulling your shots for one particular player as he does the same for you. Or you will team up briefly to take out someone who is giving both of you trouble. Other times you just find yourself outright running with a partner and it doesn't usually take long for others to notice and to team up against you. These sort of alliances (and breaking of) help to add an ebb and flow to the game yet you never see such happen with the bots. I think that implementing such a system along with a revenge and grudge system would help to make bot matches more authentic and to create more of an emotional connection between you and them (you might start seeking out a certain bot and trying to team up with him against another bot).

I believe that communication is an area of AI that is still in its infancy. Sure we can issue orders and in a few games the bots will give you a little feedback about what they are doing but this is a relatively unexplored area. Personally, unless I am directing a mindless robot army I never want to be issuing orders. Generally, in multiplayer games you rarely tell someone what to do, instead you relate to them what you want to do and they either follow you or not. At the most you can make suggestions. Whether or not that person listens to depends on how well they know you and how experienced they know you are. People who don't know you will usually only listen if they can look up at your score and see that you do know what you are talking about. A leader is usually just someone who is providing a little bit of organization. He doesn't say, "You three go down that hall while you four follow me." Orders such as that are generally ignored or argued with. Instead a leader will simply say things like, "Rush the tunnel!" or "Going for objective B" and if he is respected most of the team will follow along and feel pretty good doing so.

Communication in games tends to fall into two categories. The first is the chatting that takes place somewhat outside the gameplay. This is stuff like smack talk, taunts, jokes, observations, and just general stuff. It is often related to the game but usually takes place in the pauses of the gameplay. This type of chatting is where the "personality" of the AI can be strengthened. This can be done with something as simple as a smack talk generator or random quotes up to a lot more subtle context-based comments (much harder). The other type of communication is directly tied to the gameplay. This is when players coordinate their efforts and relay useful information to each other. I find it interesting that the AI often coordinate in their gameplay but that it is almost always a silent coordination that leaves the players out. Exposing their coordination would be fairly easy to do and would allow the players to work with the bots. The hard part would be to have the bots respond to the players coordinational conversations but I believe that this can be accomplished as well since most of this data will fall within a fairly narrow range.

Well, I believe I have summarized most of the points I made and the things I learned in the irc chat so I will leave it to others for now. I'm sure I will have more to say after I finish re-reading the logs for that chat.