Ian has been working a lot on AI recently – we talked about it quite a bit on this week’s podcast but here’s a write up and a video to give you a bit more of an insight…(watch in HD for less grunge).
Perhaps the key question when working on AI for a strategic game is whether or not it needs to display human-like qualities. Do you want the sensation of a powerful, overarching agent which is pushing its pawns around in the same manner as you, or perhaps the more immersive feeling that each unit is an individual in its own right? Mick West has a great article here about how to incorporate what he calls “intelligent mistakes” into game AI. This idea revolves around the computer knowing what makes for an interesting situation, and acting to promote this, rather than simply trying to win or lose. It will throw itself under a bus in order to add some colour.
In Frozen Synapse, we’re aiming primarily for the human element: an interesting set of ersatz players who are there whenever you want them, and who have compelling interactions with you across the playfield. The basics are in place, but soon we’ll be working on refining the system and injecting some personality.
AI is such a huge challenge because it is completely different for every game you do. Obviously there will be generic things like pathfinding involved, but you’re essentially writing code to interface with the uniquely baroque mess you’ve created every time.
That’s one of the reasons our AI is currently finding things a little challenging: it takes about five minutes to take a turn, where it should be taking a maximum of 30 seconds. Ian has to face the tough balancing act where, on one side, he has the intelligence factor, and on the other he has practicality. The vital crux here is: “How stupid can the computer afford to be?”
Unfortunately, intelligence matters in Synapse. If you mess up a single turn, it can have a bad effect. We can’t have the AI sending its harvester to the bottom of the map willy nilly (that’s a CnC analogy, there are no harvesters in Synapse…probably our first mistake…), it has to behave in a sane and reasonable manner. That’s quite a challenge for a computer.
I’ve always preferred games where the AI manifests itself in one of two ways. One: the state-shifts seem hidden; the AI appears to flow like a real player. Arguably, the AI in Unreal Tournament displays this quality. Two: the state-shifts are part of the game’s grammar (the patrol-ATTACK-frantic search-patrol of Deus Ex and Thief). This kind of game requires the first approach, and we’ll see how close we get to achieving that.
Personally, I’m excited to see the eventual composition of the single player toolkit. I’m really hoping we have the opportunity to put some fantastic content on top of a robust AI structure. These are pioneering times, but that’s what this is all about.
Please check out our blog and podcast at Visitingthevillage.com