University of Waterloo
Lingheng Meng is a PhD student, co-supervised by Prof. Dana Kulić and Prof. Rob Gorbet, in the Electrical and Computer Engineering faculty at the University of Waterloo. He is passionate about applying Artificial Intelligence technology to Human-Robot Interaction. Specifically, Lingheng applied Deep Reinforcement Learning to the Aegis and Noosphere testbed at the Royal Ontario Museum with manually designed reward functions, measuring the engagement of visitors, based on infra-red sensors. The promising results showed that higher likeability could be achieved by the proposed approach relative to pre-scripted behavior. Encouraged by this work, Lingheng is now working on developing a Preference Learning algorithm to reduce the reliance of manually designed reward functions, wherein the reward function will be learned from preferences provided by experts or novice visitors. Lingheng has also worked on developing a Unity-based simulator to test code and potentially pre-train learning agents with the help of colleagues from Philip Beesley Studio Inc.
Learning to Engage With InteractiveSystem: Embedding Reinforcement Learning into the Physical World
This research focuses on applying Machine Learning to a novel interactive robot/system, i.e. Living Architecture System (LAS), to understand how to enable such a robot to generate engaging behavior automatically. The research is approached in a step-by-step manner where Reinforcement Learning is introduced to LAS, but with manually defined reward function based on low-cost sensors and engineered parameterized action space. Lingheng is now working on implementing a reward function through user-provided preferences, which presents the potential of eliminating the need for handcrafted reward functions. Due to the complexity of the unconstrained field study, other methods are being investigated to solve the Partial Observable and non-1st-order Markov Decision Process.