Speaker: Athirai Aravazahi Irissappane

Date/Time/Venue: Feb 8, 12:00 - 1:00 PM, SIT 001

Title: Decision Making in E-Markets Via POMDPs

Abstract:  In open systems (dynamic, uncertain, insecure), users often need to rely on the abilities, competencies and knowledge of others to fulfill their own objectives. Making decisions about whom to interact with becomes cumbersome in these scenarios as the other users may be self-interested, diverse or deceptive. Further, these decisions should be made in an optimal manner, as every user strives towards maximizing its utility in the system. E-commerce, Crowd-Sourcing, Robotics and Healthcare are some of the domains where this issue of optimal decision-making under uncertainty has garnered attention.
In this talk, I will describe how we use Partially Observable Markov Decision Processes (POMDPs) to address the problem of deciding a trustworthy interaction partner in e-markets. A technique to improve the scalability of POMDPs called Mixture of POMDP Experts (MOPE) will also be presented. MOPE exploits the inherent structure of trust-based domains by aggregating the solutions of smaller sub-POMDPs. Experiments show that using MOPE one can scale to millions of states and thousands of actions.

Speaker's Bio:
Athirai Aravazahi Irissappane is a Machine Learning Engineer at Dimensional Mechanics, Seattle, USA. She is also an Adjunct Lecturer at the University of Washington, USA. She obtained her Ph.D. from Nanyang Technological University, Singapore, where she also received an Honorable Mention for her outstanding thesis. Her research interests are in Machine Learning and Artificial Intelligence focusing on Fraud Detection, Planning algorithms (POMDPs) for decision making in uncertain environments, Deep Learning for image recognition and time series analysis, etc. For further details please visit https://sites.google.com/view/athirai/