Curiously Adaptive Growing Intelligent Neural Gas: A Hybrid Approach to Category-Based Intrinsic Motivation

Intrinsic motivation, that is, internal rewards that reinforce certain behaviors in organisms, is an integral part of human development. This idea can be translated to the field of developmental robotics as a way to implement autonomous learning. This paper introduces the Curiously Adaptive Growing Intelligent Neural Gas (CAGING) algorithm, a hybrid system that builds on previous implementations to create a more robust, adaptable model. The CAGING algorithm categorizes the environment and utilizes these categories to make predictions about the robot’s sensory data. The progress that can be made by exploring a particular category acts as intrinsic motivation for the robot. The model drives the robot to maximize its learning progress by attempting to learn about novel yet predictable situations. The experiment presented in this paper, “AIBO Television,” demonstrates the effectiveness of the CAGING algorithm. In this experiment, a physical robot is placed in front of a computer monitor that displays a graphics program with three circles, each moving in a distinct way. Experimental results show that the robot focuses on the circles whose trajectories can be predicted and ignores circles with unlearnable movements. The system, while more autonomous and robust than other implementations, still suffers from some of the same problems with real world sensory data that previous systems exhibit. We conclude that the CAGING algorithm is an effective combination of existing intrinsic motivation systems.