Seminars

View all Seminars  |  Download ICal for this event

Probability Flows on Factor Graphs for Intelligent Path Planning

Series: CSA Distinguished Lecture

Speaker: Krishna R. Pattipati, Distinguished Professor Emeritus and Collins Aerospace Chair Professor of Systems Engineering, Department of Electrical and Computer Engineering, University of Connecticut

Date/Time: Sep 15 16:00:00

Location: CSA Seminar Hall (Room No. 254, First Floor)

Abstract:
This work was motivated by multi-objective, multi-agent and multi-scenario agile planning problems arising in ship routing. Using approximate dynamic programming and adaptive multi-objective A* with utopian heuristics, we have been able to solve multi-objective ship routing problems with as many as 14 objectives in <30 seconds within 1% of the optimal Pareto front. This approach extends naturally to multi-scenario and multi-agent settings with pre-specified priorities among agents. The questions we were interested in addressing were: Is there a unified approach for this problem? This is the focus of this talk.


Use of inference techniques on probabilistic graphical models to solve control/planning problems is an emerging paradigm that falls under the rubric of Active Inference, Control as Inference and Planning as Inference. We start by posing the inference problem on a Markov Decision Process (MDP) graph and show how the proposed approach, presented both in probability space and in log space, provides a very general framework that includes the sum-product, the max-product, dynamic programming and mixed reward/entropy criteria-based algorithms by simply changing the message composition rules. The framework also expands algorithmic design options for smoother or sharper policy distributions. It also extends naturally to a multi-agent system, wherein each agent follows an MDP and bases its decisions on its current knowledge and future predictions about itself and the intentions of others. In the context of multiple objectives, belief propagation dynamically generates a probabilistic Pareto front and propagates it as a forward flow distribution. We provide a comparison through simulations, first on a synthetic small grid with a single goal and obstacles, and then on a grid extrapolated from a real-world scene with multiple goals and a semantic map. Non trivial solutions and behaviors are observed in the presence of conflicting paths and objectives.

Speaker Bio:
Krishna R. Pattipati is the Distinguished Professor Emeritus and the Collins Aerospace Chair Professor of Systems Engineering in the Department of Electrical and Computer Engineering at the University of Connecticut, Storrs, CT, USA. Prof. Pattipatis research activities are in the areas of proactive decision support, autonomy, and optimization-based learning and inference. A common theme among these applications is that they are characterized by a great deal of uncertainty, complexity, and computational intractability. Prof. Pattipati received the Centennial Key to the Future award in 1984 from the IEEE Systems, Man and Cybernetics (SMC) Society, and was elected a Fellow of the IEEE in 1995 for his contributions to discrete-optimization algorithms for large-scale systems and team decision-making. He received the Andrew P. Sage award for the Best SMC Transactions Paper for 1999, Barry Carlton award for the Best Aerospace and Electronic Systems (AES) Transactions Paper for 2000, the 2002 and 2008 NASA Space Act Awards, the 2003 AAUP Research Excellence Award and the 2005 School of Engineering Teaching Excellence Award at the University of Connecticut. Prof. Pattipati served as Editor-in-Chief of the IEEE Transactions on SMC-Cybernetics (Part B) during 1998-2001.

Host Faculty: Prof. Y Narahari