Exploiting interactive genetic algorithms for creative humanoid dancing

The paper discusses an approach aimed at endowing a cognitive architecture with artificial creativity capabilities in order to make a humanoid able to dance in a pleasant manner. The robot associates movements to music perception creating an aesthetically valuable dance by using a Hidden Markov Model with a nonclassical approach. Two matrices mainly influence the model: a Transition matrix TM, and an Emission Matrix EM. The TM matrix rules the transition between two subsequent movements. The EM matrix constitutes the link between a set of movements and the perceived music features. In order to compute the EM matrix, we exploit a genetic algorithm approach. The approach makes use of two kinds of fitness functions. The first one is an internal evaluation fitness that allows the robot to autonomously learn the association between music and movements. The second one depends on the interaction with a human teacher, leading to the determination of different dance styles, which constitute the robot repertoire. The experimental part discusses the effects on the creativity of different distances to compute fitness.