Affiliations: Control and Intelligent Processing Centre of Excellence, School of Electrical and Computer Eng, University of Tehran, Tehran, Iran | School of Cognitive Sciences, IPM, Tehran, Iran | Autonomous Systems Laborartory, ETH Zurich, Switzerland
Note: [] Corresponding author. Maryam S. Mirian, Control and Intelligent Processing Centre of Excellence, School of Electrical and Computer Eng, University of Tehran, Tehran, Iran. E-mails: mmirian@ut.ac.ir, araabi@ut.ac.ir (Babak N. Araabi), mnili@ut.ac.ir (Majid Nili Ahmadabadi), rsiegwart@ethz.ch (Roland R. Siegwart).
Abstract: Rapid increase in the size and complexity of sensory systems demands for attention control in real world robotic tasks. However, attention control and the task are often highly interlaced which demands for interactive learning. In this paper, a framework called METAL (mixture-of-experts task and attention learning) is proposed to cope with this complex learning problem. METAL consists of three consecutive learning phases, where the first two phases provide an initial knowledge about the task, while in the third phase the attention control is learned concurrently with the task. The mind of the robot is composed of a set of tiny agents learning and acting in parallel in addition to an attention control learning (ACL) agent. Each tiny agent provides the ACL agent with some partial knowledge about the task in the form of its decision preference- called policy as well. The ACL agent in the third phase learns how to make the final decision by attending the least possible number of tiny agents. It acts on a continuous decision space which gives METAL the ability to integrate different sources of knowledge with ease. A Bayesian continuous RL method is utilized at both levels of learning on perceptual and decision spaces. Implementation of METAL on an E-puck robot in a miniature highway driving task along with farther simulation studies in Webots™ environment verify the applicability and effectiveness of the proposed framework, where a smooth driving behavior is shaped. It is also shown that even though the robot has learned to discard some sensory data, probability of raising aliasing in the decision space is very low, which means that the robot can learn the task as well as attention control simultaneously.
Keywords: Attention control learning, decision space, perceptual space, bayesian continuous RL, learning to drive
