PLURELEARN · Plural Reinforcement Learning

We propose a new paradigm for learning in complex high-dimensions dynamic environments. Our goal is to develop algorithms, theory, and applications that use plurality of learning approaches and models in a synergetic way. Our paradigm considers the task of learning a control policy by combining trial and error in the style of reinforcement learning with learning from a competent teacher whose interaction with the environment can be observed. Instead of using the teacher for imitation, our paradigm is focused on learning good representations of the world-model. We consider four specific issues in the new paradigm: (i) The usage of iteration and reiteration between learning from a teacher and reinforcement learning. (ii) Learning representation and structure from the teacher. (iii) Optimizing policies based on learned representations and reasoning about model uncertainty. (iv) Learning sub-strategies from a teacher and when and how to use them. We will develop algorithms and theory pertaining to the new paradigm and will apply it in two challenging domains: a fighter jet simulator and a network operating center simulator.