[TISEA 1992] Paper: Patti Maes – Modelling Actions Selection and Learning in a Software Pet


For the past couple of years I have been developing models of artificial creatures. These
els are not only interesting from a biology point of view (they can be used to test theories about animal behaviour), but also from an entertainment point of view, since they can be used to build autonomous creatures with which a human can have interesting interactions. In particular, I developed an algorithm for motivational compe-tition and selection of behaviours in an artificial creature. In more recent years I added a learn-ing component to this algorithm which allows the creature to change its behaviour selection based on experience and environ-ment feedback. I have built several AI creatures (robotic and simulated ones) which illustrate and test these algorithms. I will show videos of some of these creatures and most likely give a live demo. I will focus on a recent experiment in which we mod-elled a live “dog” which lives in a simulated two-dimensional environment. A user can interact with the dog: change its motivational levels, teach it some new behaviours by giving positive and negative feedback and observing the resulting (and sometimes surprising) behaviour.

  • Patti Maes (Belgium/USA), MIT MediaLab, USA