tw@aipna.ed.ac.uk (Toby Walsh) (04/16/91)
Building Embedded Agents Leslie Pack Kaelbling Stanford University and Teleos Research It is very difficult to build agents that operate continuously in dynamic, unpredictable worlds. There are two ways that this task can be made easier: by designing new programming abstractions, languages, and tools to simplify the problem for human programmers; and to develop learning algorithms that move some of the burden from the programmer to the agents themselves. In this talk, I will present work in each of these areas. In programming agents, there is a strong tension between using low-level formalisms, which are efficient to execute but difficult for humans to use, and using higher-level declarative formalisms, which simplify the programming task for humans, but are often computationally intractable to execute. I will describe Gapps, a language for programming embedded agents at the declarative level in terms of goals of achievement and maintenance that compiles into efficiently executable programs with bounded reaction time. Gapps has been used effectively in robot programming projects. Good programming formalisms are not enough. Many times the programmer does not know everything about the domain in which the agent will be working, so the agent must learn from its experience with the world. The area of reinforcement learning has received a great deal of attention from statisticians and learning-automata theorists, but has not been a focus of artificial intelligence research. I will give a brief discussion of the formal setting and some important problems in reinforcement learning, then go on to present new work on efficient algorithms for reinforcement learning. Finally, I will speculate on ways to integrate these two approaches in order to exploit the strengths of both. ---------------------------------------------------------------------- For more information about this or future seminars please contact Toby Walsh <T.Walsh@ed.ac.uk> -- Toby Walsh, JANET: Toby_Walsh@uk.ac.edinburgh Dept of Artificial Intelligence, ARPA: Toby_Walsh%uk.ac.ed@nsfnet-relay.ac.uk 80 South Bridge, UUCP: ...!ukc!ed.ac.uk!Toby_Walsh Edinburgh EH1 1HN TEL: +44 31 650 2725