Shaping: A multiple contingencies analysis and its relevance to behaviour-based robotics

Abstract

Operant shaping is a technique used by psychologists to train new behaviours in animals and is also of interest to roboticists and artefactual modellers. This review argues that an understanding of the nature of shaping in animals will facilitate the use of this training method by roboticists, and other designers of artificial adaptive systems, who wish to adopt ideas and concepts from animal psychology and ethology. To do this, the paper reviews a number ofanimal models of operant shaping. It identifies issues that are relevant to the successful application of this technique in both animals and autonomous agents. These issues include the question of how to identify and assess the behaviours subject to shaping, the nature of the learning contingencies which are active during operant shaping, and the issue of whether shaping generates new and original functionalities or, alternatively, simply acts to reorganize the existing behavioural repertoire of the animal or robot. A number of robotic applications of shaping techniques are considered. The similarities between ontogenetic shaping and the phylogenetic shaping techniques employed in evolutionary learning are discussed. The paper concludes by arguing that an understanding of ontogenetic shaping, achieved through the study of animals, will facilitate the design of robots that must function in real world situations.

Keywords:

• Shaping
• Successive Approximations
• Multiple Contingencies
• Evolutionary Learning
• Fitness Functions
• Critical Situations