Speaker: Ankit Anand

Time: February 25th, 3 pm

Title: Exploiting Symmetries in Sequential Decision Making under Uncertainty

Abstract: The problem of sequential decision making under uncertainty, often modeled as a Markov Decision Process(MDP) is an important problem in Artificial Intelligence community. Traditional MDP solvers operate in flat state space and don't scale well in large state and action spaces. A lot of real world domains have exponential number of states in terms of representation but many of these states and actions are symmetric to each other. In this talk, we focus on exploiting symmetry in these domains to make state of art algorithms more efficient and scalable. Our recently proposed ASAP framework: "Abstraction of State-Action Pairs" extends and unifies past work on domain symmetries by holistically aggregating "state-action pairs" in addition to states. ASAP uncovers a much larger number of symmetries in a given domain. We also discuss a batch algorithm to use ASAP framework in Monte Carlo Tree Search framework specifically UCT. We also discuss our recently accepted work of "OGA-UCT: On-the-Go Abstractions in UCT" which is the first algorithm to compute symmetries on-the go (while building tree) in UCT.  

Bio:  Ankit Anand is a graduate student in IIT Delhi working with Dr.Parag Singla and Dr. Mausam. His research interest lies in Artificial Intelligence and Machine Learning where he focuses on inference in Probabilistic Graphical Models and Planning. Prior to joining here, he obtained his masters from Indian Institute of Science, Bangalore where he worked with Prof S.K.Nandy in Cloud Systems. He obtained his Bachelors from Punjab Engineering College, Chandigarh. He also had a short tenure at Samsung India Research Labs in Noida.