Self-Organizing Maps with Eligibility Traces: Unsupervised Control-Learning in Autonomous Robotic Systems

This paper presents the application of a connectionist control-learning system to autonomous mini-robots. These robots are specially designed and equipped for learning easily deened yet challenging control tasks. The system must learn despite noisy, low-grain sensory input and uncertain interactions between motor commands and eeects in the world. The system's design is severely constrained by the computing power and memory available on board the mini-robots and the on-board training time is limited by the short life of the battery. The connectionist system proposed ts into the Self-Organizing Neural Network with Eligibility Traces (SONNET) paradigm proposed by Hougen 7]. SONNET systems are capable of unsupervised learning of input space distribution (partitioning) and output responses in temporal domains. These systems are based on the well-known, self-organizing topological feature maps of Kohonen. They are augmented for response learning by the addition of eligibility traces. This combination of features allows these systems to solve diicult temporal credit-assignment problems rapidly through the use of information sharing in input and output neighborhoods.