Graph-Based Bandit Problems
We exploit graph signal processing to derive efficient optimization strategies for large scale problems. The key intuition is to learn in a sparse but representative domain. Possible applications are: optimal placement of heating sources in sensor networks, optimal ads in social networks, orrecommendations in high-dimensional problems.
Structural Reinforcement Learning
Classical reinforcement learning problems suffers from the well-known curse of dimensionality, leading to very slow learning problem in high-dimensional search spaces. We target to overcome this limitation is to infer the structure/geometry of the problem in the learning algorithm in the case of high-dimensional states. Possible applications are online control of grid networks or traffic networks.
A major challenge for the next decade is to design virtual, augmented, and mixed reality systems (VR at large) for real-world use cases such as healthcare, creative technology, e-education, and high-risk missions. This requires VR systems to operate at scale, in a personalised manner, remaining bandwidth-tolerant whilst meeting quality and latency criteria. This can be accomplished only by a fundamental revolution of the coding/streaming/rendering chain that has to put the interactive user at the heart of the system rather than at the end of the chain.
Do we understand how people interact with technology? Can we model users’ behaviour in VR systems, inferring the correlation between users’ interactivity and the VR content?