A curious design agent: A computational model of novelty-seeking behaviour in design

Abstract

This paper presents a "curious design agent", i.e. an agent that uses the search for novel designs to guide its design actions. A computational model of curiosity based on a process called novelty detection is presented. The behaviour of the computational model is illustrated with a curious design agent searching the space of two-dimensional patterns generated by a simulated Spirograph is reported. 1. A Computational Model of Curiosity Curiosity is the motivation to discover new knowledge when faced with an unfamiliar situation (Berlyne, 1971). It promotes self-directed learning by rewarding behaviour that results in the assimilation of new knowledge. Curiosity can be used to guide the search and exploration of unfamiliar design spaces to find new knowledge with the goal of gaining a better understanding of a non-routine design task. Curious search can be used to guide problem solving to find interesting design solutions. Curious exploration can be used to guide problem finding to discover interesting design problems. In this paper we concentrate on the role of curiosity in the search for interesting design solutions. The curious design agent presented here uses a computational process called novelty detection to determine its level of interest in designs. The curious design agent uses the detection of novelty to guide its search during the course of a design session. Determining interestingness based on novelty depends upon the knowledge of the agent and its computational abilities; things are boring if either too much or too little is known about them (Schmidhuber, 1997). A neural network called a self-organising map or SOM (Kohonen, 1995) is used in the following experiment to implement a simple novelty detector. SOMs project vectors in the high dimensional space onto a hyperplane, typically two-dimensional. The projection provides an ordered map of the high dimensional space with similar vectors clustered together into neighbourhoods.