Self-motivated Learning Agent-Skill-development in a Growing Network Mediated by Pleasure and Tensions

摘要

We present a novel method of machine learning toward autonomous developmental systems. The method is based on a growing neural network that initially produces senseless signals but later associates rewarding signals and quasi-rewarding signals with recent perceptions and motor activities and, based on these associations, incorporates new cells and creates new connections, which results in more structured output patterns. The rewarding signals are produced in a device called “pleasure center”, while the quasi-rewarding signals (that represent pleasure expectation) are generated by the network itself. The network was tested using a simulated mobile robot equipped with a pair of motors, a speaker, a set of touch sensors, and a camera. Despite a lack of innate wiring for any purposeful behavior, the robot developed from scratch, without any external guidance (except hardwired perception-pleasure patterns), a set of perception-reaction patterns. The emerging patterns include obstacle avoidance, vocalization of interest, and approaching an object of interest, which are fundamental for creatures and usually handcrafted in traditional robotic systems.