Friday, November 16, 2007

Catan: Worst AI evar?

So, one of my coworkers made the bold claim that the AI for Catan! (XBLA) was the worst in the universe. I'd heard some complaints about the AI before, so I asked him to explain why he felt the way he did in order to get to the bottom of the issue.

You see, I worked with the co-developers of Catan's AI... Klaus Teuber (the creator of Settlers of Catan) and Brian Reynolds ("game god" behind Alpha Centauri, Civ II, Rise of Nations). In other words, two acknowledged heavy-weights were involved in the creation of Catan's brain.

Reading into the reasoning of my co-worker, it seemed like at least part of the frustration was that the AI couldn't explain what it was doing -- and part was because the AI couldn't be reasoned with if it seemed to be acting irrationally. This is something that can be solved naturally with human friends (through chatting) but is harder to emulate with computers.

One way we attempted to address this problem was to add emotes (tickles) to help educate users about what the computer was thinking. However, while some things were easy to figure out (computer smashes the piggy bank means he doesn't have the resource you're looking for) others were much more subtle (computer flashes "not with you, now" when he is competing with you to settle an open vertex).

I wonder if one of the other things we may have done "wrong" for the SP experience is that we optimized AI play for filling in dropped players in an MP game. We specifically didn't want them to do stupid trades with the human leader as it would make the end game not fun for the other human players. The fact that even though all 3 AI players are operating independently and according to their own interests doesn't matter. What it looks like is that all 3 AI players are ganging up on the human in an arbitrary and unfun way.

Of course, when you find yourself in a game with 3 expert human players (as I often did while working at BHG) the end game will end up with a 3 against 1 mentality as no one will trade with the leader unless it guarantees them a victory that turn. We tried to mitigate against this in the Easy and Normal difficulty settings, but it's obvious that we didn't do enough. And to the player, this can seem like a broken experience.

The other thing that I find infuriating (but less so now that I know about it) is that sometimes computer players have "crises" depending on their own personality preferences and how the board is set up. This means that they will panic and be basically unwilling to engage in any behavior that won't let them make a specific play (and unwilling to assist anyone who is competing for the same play). Tournament level players (like Brian) would see the AI's inflexibility and read the board and understand the dilemma. Normal human players (like me) will only see a completely unsocial AI that seems unwilling to trade. Note that it's not the "crises" that I find infuriating (AI should have goals)... It's the fact that I don't know why the computer is behaving so anti socially.

I think I'm going to have to think much more deeply about AI as a user experience. How do we have computer operated agents that work in believable, understandable, and entertaining ways?

No comments: