The latest research developments to enable robots and other “automated planning agents” to maximize their on-board computational powers will be the focus of a Lawrence University Science Hall Colloquium.
Kurt Krebsbach, associate professor of computer science at Lawrence presents, “Planning to Plan: Deliberation Scheduling using GSMDPs,” Thursday, April 13 at 4:15 p.m. in Science Hall Room 102 The lecture is free and open to the public.
With limited computational resources such as time, memory and partial information, robots in realistic (over-constrained) situations are unable to produce the perfect sequence of actions because the “deliberation” required to do so is unavailable. The problem of deliberation is magnified when acting and planning occur concurrently because satisfactory plans must be constructed in time to be executed.
Just as people in the real world are forced to “think about what to think about” all the time, Krebsbach says researchers are turning to meta-planning, or “planning to plan,” to help robots determine which planning activities are worthwhile given the constraints of the situation at hand.
In his presentation, Krebsbach will discuss how the problem of deliberation scheduling is being addressed by “decision-theoretic approaches based on recent advances in Generalized Semi-Markov Decision Processes (GSMDPs).” This first-ever application of GSMDPs to the problem of deliberation scheduling will allow computer scientists to more accurately model domains in which planning and execution are concurrent, plan-improvement actions have uncertain outcomes and durations and events, such as threats, occur randomly.
A specialist in artificial intelligence and automated planning, Krebsbach joined the Lawrence faculty in 2002. A 1985 Lawrence graduate, he earned bachelor’s degrees in both music and mathematics/computer science. He holds an M.S. and Ph.D. in computer science from the University of Minnesota.