The Sequential Centipede Applet

This applet plays a variant of the well-known risk assessment game called Centipede. This is played between two players. When the game starts, there is a $100 pot of money. The first person to hit their button wins the entire pot. However, for every second that a player delays, the pot doubles. Therefore, a successful strategy involves waiting long enough to let the pot grow to a substantial size, but not so long where the opponent will hit the button first.

This project uses three different neural networks, in an effort to understand which of the three recurrent types would best capture human-like behaviour. Since neural networks are a trained system, several games between human opponents were used as the training data. Once the neural networks were trained, they were each played against the others, as well as a random number generator.

The original proposal for the experiment is available here.