Deep Reinforcement Learning with Subsumption Architecture for Embodied Agents

Abstract: Deep Reinforcement Learning methods have gained a lot of attention in recent years due to their ability to learn complex behaviors based only on high-level sensory input and simple rewards as the feedback. Deep Q-Networks (DQN) could be a well-fitted solution for controlling embodied agents such as mobile robots, which need to interact with the environment using their senses and actuators. The subsumption architecture is a classic method for controlling embodied agents using parallel, loosely coupled processes. The subsumption architecture includes a hierarchy of layers – each layer is controlling different aspect of the agent’s behavior and may inhibit or suppress actions of other layers. In my research I am showing how DQNs may be used with the subsumption architecture and thus, how such system may learn to control an embodied agent. Speaker Bio: Tomasz Tajmajer is a PhD student of computer science in the Faculty of Mathematics, Informatics and Mechanics at the University of Warsaw where he works under the supervision of Dr Piotr Wasilewski. He received a MSc in telecommunications engineering from the Warsaw University of Technology where he also participated in several research projects related to Wireless Sensor Networks and Smart Cities. Currently he is interested in embodied artificial intelligence and his area of research includes Deep Neural Networks and their applications to cognitive science and robotics. He is also a member of the Mars Rover Team at the University of Warsaw.

More information at: https://www.dal.ca/news/events/2016/12/13/deep_reinforcement_learning_with_subsumption_architecture_for_embodied_agents_.html