Basic Language Learning in Artificial Animals

We explore a general architecture for artificial animals, or animats, that develops over time. The architecture combines reinforcement learning, dynamic concept formation, and homeostatic decision-making aimed at need satisfaction. We show that this architecture, which contains no ad hoc features for language processing, is capable of basic language learning of three kinds: (i) learning to reproduce phonemes that are perceived in the environment via motor babbling; (ii) learning to reproduce sequences of phonemes corresponding to spoken words perceived in the environment; and (iii) learning to ground the semantics of spoken words in sensory experience by associating spoken words (e.g. the word "cold") to sensory experience (e.g. the activity of a sensor for cold temperature) and vice versa.