Efficient Learning of Multiple Degree-of-Freedom Control Problems with Quasi-independent Q-agents

We propose a new architecture in which autonomous or quasi-independent agents cooperate to solve control problems in large action spaces. Action space is decomposed according to its degrees of freedom, and each agent uses the Q-learning algorithm to control actions along a single action space dimension. The technique outperforms a single agent in the control of a small, abstract speech articulation model. It also learns to perform some simple articulatory tasks in controlling a realistic vocal tract model with 10 degrees of freedom.