Social Reinforcement in Artificial Prelinguistic Development: A Study Using Intrinsically Motivated Exploration Architectures

This paper introduces an intrinsically motivated sensorimotor exploration architecture which considers social reinforcement and motor constraint awareness. The main objective is to study the influence of social interactions during artificial early prelinguistic development. We argue that this architecture contributes to explain development from voiceless to sequence of vowels vocalizations. A cognitive developmental perspective is considered emphasizing embodied cognition and sensorimotor exploratory behaviors. For a new-born agent, motor constraints are unknown. However, the agent is endowed with a somatosensory system that indicates if a motor configuration was reached or not. This information is used to model and predict constraint violations. Furthermore, the architecture considers imitative behaviors that constrain the search space during exploration. Interaction occurs when the learner sensory production is similar to a sensory unit relevant to communication. In that case, the instructor perceives this similitude and reformulates with the relevant sensory unit. When the learner perceives an utterance by the instructor, it attempts to imitate it. Two systems are considered for experimentation: 1) a toy example and 2) a simulated vocal tract. In general, our results suggest that constraint awareness and social reinforcement contribute to achieve less redundant exploration, lower exploration and evaluation errors, and a clearer picture of developmental transitions.