Rock, Paper, Scissors redux, now with AI

The New York Times recently posted about an AI-backed Rock-Paper-Scissors game. I have written about it before and some have disagreed with me.

The AI version learns from past games and uses that data to predict future moves. There are two modes: “Novice” and “Veteran”. The former only learns from the rounds you’ve played with it, the latter draws from all previous games that have been played with it.

If you are someone, like my naysayer in the link above, that buys into the illusion that you can adopt any kind of real strategy in Rock-Paper-Scissors, the AI will smoke you. If you even pay attention to what the computer (or any opponent) does at all, you are buying into the lie and will become predictable.

As that page mentions, “[a] truly random game of rock-paper-scissors would result in a statistical tie with each player winning, tying and losing one-third of the time.” This is true assuming that both players are playing randomly, ignorant of one another’s choices. If one player adopts a random strategy (“plays irrationally”) while the other one plays with a strategy (“rationally”), the player with no strategy is unaffected, but the player with the strategy will be at a slight disadvantage.

A given choice in RPS is identical to the other possible choices — each one can win, lose, or tie the round with equal probability. The choices have no inherent value until the opponent’s choice is revealed. When both players play rationally, they attempt to out-think what their opponent will pick, and assign a weight value to the available choices based on that prediction. The thing to consider here is that the weight assigned is entirely based on the assumption that the opponent is selecting something on a rational basis. By selecting your choice randomly, you completely negate any advantage the other player would have gained through their strategy; not only that, but there is also a subtle cumulative penalty that they incur for as long as they believe you are still acting rationally.

When playing Texas Hold’em with some friends, a few years, ago, I realized that I was outmatched on skill and experience, so I instead adopted a completely irrational strategy – I would bet erratically, irrespective of what I was actually holding (unless I had a REALLY strong hand) and I ended up winning. Granted, my friends aren’t professionals, and that strategy would probably fail against real pros.

The fundamental difference between Poker and RPS is that RPS has no memory of previous rounds, unlike Poker. Each round in RPS may as well be a roll of the dice, a completely separate statisical event. You can play completely randomly and are not penalized cumulatively, like you would be with chips in Poker.

The screen-cap above is a small sample run I did against the computer, using a random number generator to make my choices. In a range of 1-3, 1 is Rock, 2 is Paper, 3 is Scissors. No matter what the computer plays, stick with the random protocol. You might as well just hide half of the screen, so that you aren’t tempted to make a “strategic choice.” After 40 rounds of random choice, my final record was 16W, 11T and 13L. Admittedly, the sample size isn’t very large, but I think it at least provides some good support to my argument: regardless of whether or not your opponent is using any kind of strategy, a random play is strictly better, and the more rationally they try to play the worse they do.