Scientifica 2017: rock-paper-scissors

Nicolai Meinshausen (NM) and Martin Mächler (MM) from the Seminar for Statistics have developed what they call the "Statistics Machine" and are using it to engage visitors to their stand in a game of rock-paper-scissors. In this interview, they talk about the reasons for choosing the game and how they are using statistical methods to predict human playing behaviour.

What inspired you to set up your own stand at Scientifica?

NM: We felt inspired by the motto of this year's event – "What data can reveal". It really sums up what statistics is all about – so, as statisticians, we felt that we had to get involved somehow.

MM: The key word here is "reveal". Statistics doesn't mean organising data – it means using it to draw conclusions about what is happening in the real world.

Why did you choose the game of rock-paper-scissors?

NM: The game needs to be something that people will pick up quickly and is easy to explain. It also has to be something that will open the door to conversations with people about things like learning from data and using data to make predictions.

In the game, you assume that people's actions are not random, but instead follow a certain pattern. How do you know that?

NM: From observations of human behaviour, we have known for quite some time now that it is generally very difficult for people to generate random sequences.

MM: Let's say that someone is asked to imagine tossing a coin a hundred times and make a note of the results each time, as opposed to tossing a real coin. Experts can immediately tell the difference between the real and imagined scenarios.

Does that mean people are unable to think in random sequences?

MM: Yes – people think that chance works in a different way than it actually does. In a coin toss, for example, people believe that the chance element lies in the potential outcome changing as many times as it possibly can. In reality, the potential outcomes in a chance situation change much less than this.

So if you were given the task of noting down a random sequence, would you be better at doing that if you had a better understanding of chance?

MM: I don't think so. If you really want to generate something random, a computer is what you need.

To predict the moves in rock-paper-scissors, you need a data model that is able to forecast the behaviour of a human player. How did you generate this?

NM: We programmed various strategies, all of which have very simple concepts. One strategy assumes that, if a player wins, they will repeat the same move as before – but if they lose, they will do something different the next time. In another strategy, the player never repeats the move; instead, they keep changing it. Whenever the machine plays against a person, the model is then able to predict which strategies that person is using.

MM: People won't always play in the same way – they'll adapt their strategy to what is currently happening in the game, so it's changing all the time. And every person has a different way of playing.

NM: The machine analyses what the player is doing and adapts itself in line with this. We had our doctoral students play against the machine and adjusted its system using the results to ensure that it would be able to play well.

How long did that process take?

NM: We analysed a few thousand moves carried out by around ten to fifteen doctoral students, and that gave us the tools to fine-tune the system as effectively as possible. But the system also adjusts itself to every new player it encounters.

How will the game work at the stand?

NM: People will play a sequence of 25 games on a tablet. The next move will be predicted on a second screen. The person playing won't be able to see this, but the other visitors to the stand will.

Are you using these methods to analyse playing behaviour in research too?

NM: Our research essentially concentrates on creating the ability to make accurate predictions. We look at how cells react to genetic changes, for example, or how certain we can be when predicting climate change. One common factor that you find throughout data analysis work nowadays is that the few pieces of data that are valuable and useful are often buried in a vast mountain of somewhat useless data. We're always looking for the proverbial needle in a haystack.