Danny ChaiThis was a really interesting paper about how AI algorithms can break games in interesting ways. There’s one game (CoastRunners) where to maximize the score, the AI learned that you don’t actually try to win the boat race. Instead, you just circle around one section where scored targets regenerate and keep racking up the points. The post argues that this is a general AI problem of having an incorrect reward function – if you set it up wrong you get questionable, even dangerous AI behavior. But to me it also reveals that some games are just fundamentally flawed.
For some reason it reminded me of this kids’ game show I watched growing up: Fun House. We didn’t have cable so I never saw all the kids shows on Nickelodeon, including Double Dare, which I only knew by reputation. I think Fun House was syndicated. One of my mild regrets is that my sister and I never tried to get on the show, because like CoastRunners, I think it was fundamentally flawed and we could have exploited it. In the show the competition between two teams starts with questions and challenges for points. (I think my sister and I would have been at a significant disadvantage for physical challenges but at an advantage for questions.) Then the last competition before determining the winning team was a race.
The race was what was flawed. Throughout, there are stations where you can stop and get tokens (one worth 10 points, one worth 25). But winning the race was only worth 25 points. Given this setup, obviously trying to get every token along the race was way more important than actually winning. But in my memory, teams frequently – almost always – prioritized trying to win the race and would ignore tokens they dropped, quickly move on if they were too hard to get, whatever. It made no sense, but most teams did that. Maybe it was the adrenaline. Or maybe just most kids are dumb. Anyway, it drove me crazy to watch and made me wish we were on the show. We would have ruined the entertainment though, as you’d see one team racing and the other (us) casually and carefully collecting every token along the track. But not our theoretical fault – it was a fundamentally flawed game.
In general, I find it interesting which games and sports get better with analytical optimization and which get worse. (American) Football I think gets better. The things analytics says are optimal – going for it more frequently on 4th down, passing more in general – make the game more exciting and fun to watch. Baseball I think is the opposite. I applaud the rule changes they’re making in an attempt to get more action, but it’s tough because optimal baseball means less action – all strikeouts and homers. Basketball is somewhere in between. Some things are good – greater pace, less mid-range. But it’s also become a little homogenous in style and made post play relatively rare. A lot of games come down to a 3-point shooting contest and that’s not super fun.