Wednesday, Sep. 12, 2007

IEEE Computational Intelligence Society

Automating Dynamic Planning and Execution in a Partially Observable Game Model

Speaker: Jim Vaccaro, PhD Candidate

About the program: The talk centers on adapting a plan for a partially observable environment. The application discussed is a search and rescue operation for a small city, where 20000 people are stranded in over 3500 buildings after the city is flooded by either a tsunami or hurricane.  In the wake of the Katrina incident in New Orleans, more adaptive planning capabilities are needed where reaction times are so important.  The game model has over 30000 waypoints, including sky for helicopters, roads for busses, water for boats, and internal building locations where people are stranded.  Water level, people whereabouts and road blocks are not known a priori.  Rescue routes for boats, busses and helicopters were calculated a priori, but the disaster has changed the situation so that prior shortest routes may no longer exist and trees may be blocking bus routes.  The problem is where to send the boats, busses and helicopters to evacuate the city most efficiently.  In addition, boats and busses are used for rescue operations only, but helicopters can either survey damage or rescue people and the Planner needs to make these tough decisions quickly.  Through the use a tournament learning algorithm which uses evolutionary search, decisions on where to send vehicles for rescue is adaptively changed to reduce the overall rescue time.  The first half of the talk will give an overview of the approach taken and the second half will go over the results in detail.  Findings include the importance of relaying information back to base (the planner) and what state variables are important in making decisions on the fly.  Delays are incurred in travel times, searching buildings, pick ups, awaiting orders, and mishaps.  For simplicity, the delay times and travel speeds of vehicles are consistent throughout the simulation.

About the Speaker: Mr. James M. Vaccaro, PhD Candidate, UCSD and Staff Scientific Researcher, Lockheed Martin, (858) 795-1441 jim.vaccaro@lmco.com; Mr. Vaccaro has a background in reliability science, modeling and simulation and dynamic planning technology with a focus on their computational intelligence aspects such as game theory, recurrent systems, pulse-coupled neural networks, Hebbian learning, genetic algorithms, Bayesian networks, reinforcement learning, and hybrid systems.  He has worked at a variety of research institutes: Air Force Research Laboratory, UC Berkeley, University of Southern California, Institut fur Neuroinformatik, Bochum, Germany, Centre DEtudes et de Recherches, Toulouse, France and UCSD.  His current interests are in self-aware planning algorithms, where the planning algorithm improves its strategy through tournament play.  Note that Dr. Clark Guest at UCSD is a co-contributor in this work.

Time/Place: Wednesday Sep 12, 6:00 P.M. Lockheed Martin, 4770 Eastgate Mall San Diego, California 92121. Food served starting at 6:00 p.m., talk will begin sharply at 6:30. Directions and lecture background materials available at the SD CIS website. http://ewh.ieee.org/r6/san_diego/cis/

Information: Andrew Diamond (IEEE CIS San Diego Chapter Chair) (858) 509-3115, adiamond@EnvisionSystemsLLC.com