Exploration algorithm for learning of sensorimotor tasks using sampling from a weighted Gaussian Mixture

This study presents a sampling efficient algorithm of a goal-directed exploration for learning complex non-linear sensorimotor mappings. The proposed generic approach uses sampling from weighted Gaussian Mixture Models (GMs) with both positive and negative weights that is shown to be an efficient way of searching in a non-linear space with multiple local minima. The simulations were performed by training the articulatory model to learn five distinct sounds of English vowels: [a], [e], [i], [o], [u]. The results demonstrated that after 400 iterations, the algorithm generated sounds with the competence values above 82% for all 5 vowels.