From Sensorimotor Graphs to Rules: An Agent Learns from a Stream of Experience

In this paper we argue that a philosophically and psychologically grounded autonomous agent is able to learn recursive rules from basic sensorimotor input. A sensorimotor graph of the agent's environment is generated that stores and optimises beneficial motor activations in evaluated sensor space by employing temporal Hebbian learning. This results in a categorized stream of experience that feeds in a Minerva memory model which is enriched by a time line approach and integrated in the cognitive architecture Psi--including motivation and emotion. These memory traces feed seamlessly into the inductive rule acquisition device Igor2 and the resulting recursive rules are made accessible in the same memory store. A combination of cognitive theories from the 1980ies and state-of-the-art computer science thus is a plausible approach to the still prevailing symbol grounding problem.