Undead Astronauts

There’s a competition for almost everything and that includes rock, paper, scissors of course. Repeat winners (or at least repeat high placers) are common because people are not random, and there’s skill involved in predicting how opponents will react.

The NYTimes has a fun little game where you can play RPS against a veteran computer program which has a backlog of thousands of games upon which it draws statistical inferences about how people tend to play. It was unnerving playing against it for my first few games because, going on instinct alone, I lost or tied every game for 18 games (0-7-11). It was like I was a violin and the computer just plucked me as it pleased. Finally, I hunkered down and vowed to end with a better record than the damn thing. First, I realized that I had to metagame the computer. It operates by looking at what both I and the computer threw in the last few hands, whether we won or lost, and compared that to what it’s database said would be the next likely move. But that isn’t all it did. I’m pretty sure that it learned based on patterns of the individual player also.

I began winning when I consciously tried to metagame the backlog of plays. I mimicked the computer by looking at the past few rounds, decide what my gut told me to throw on my next hand based on that, then throw the hand that would beat the hand that the computer thought I would throw. This tactic took a lot of concentration and it worked pretty well up for awhile. I won 5 or 6 hands in a row at one point, and then the computer caught on. I lost 3 or 4 hands right back with my metagame strategy, so I then decided to abandon that strategy and throw exactly what my gut told me to throw and won a couple of rounds. Using a mix of strategies including metagaming, randomness (would quickly move my mouse with my eyes closed and click), and gut instinct, I finally had the winning record at 59-52-58.

My takeaway from this is that while the computer is very good, it’s still vulnerable to someone who knows how it works. It never took the initiative, as far as I could tell, in switching up strategies. It only reacted by basing it’s next move on my last few moves and it’s catalog. If a player realizes this, he can out-meta the comp by switching strategies every few hands so that the computer can’t rely on consistent play. If this is the state-of-the-art in today’s AI, I’m impressed but not overwhelmed. There’s no Skynet yet.