Files in this item



application/pdf9314845.pdf (7MB)Restricted to U of Illinois
(no description provided)PDF


Title:A Machine Learning Approach to Planning in Complex Real-World Domains
Author(s):Bennett, Scott William
Doctoral Committee Chair(s):DeJong, Gerald F.
Department / Program:Electrical Engineering
Discipline:Electrical Engineering
Degree Granting Institution:University of Illinois at Urbana-Champaign
Subject(s):Engineering, Electronics and Electrical
Artificial Intelligence
Computer Science
Abstract:Classical planning techniques have some serious problems when employed in real-world domains. In classical planning, it is assumed we know the current state of the world and can project that state through a reasonably well-defined set of actions to yield a future state. However, perfect models of the world and of operators are not possible in most domains. Consequently, discrepancies occur between the projected future state and the observed future state. In these complex domains, the success of the plan can never be guaranteed. Furthermore, an important tradeoff exists between the time spent constructing a plan and its resulting chance of success. Several approaches to these problems have been investigated, including the use of decision-theoretic methods and the incorporation of reactivity into planners. We present a new technique called permissive planning. Explicit approximations are employed in representing the world state and operators. Plans are then constructed efficiently using the approximate theory. In response to plan execution failures, plans are refined so they become less sensitive to the approximate knowledge used in their initial construction. This is achieved by tuning parameters of the plan so as to minimize the expected future deviation.
Each permissive plan has a target success rate and a degree of confidence desired in that success rate. We present a formal permissive planning algorithm which can be shown to either produce a plan with the desired success rate and degree of confidence, if possible, or otherwise to return the plan falling short of the target but with the best possible success rate. One of the downsides of this is that many examples are needed to gather the statistical evidence necessary to ensure these claims. Consequently, we propose an approximation to this algorithm which uses heuristics to determine how to refine plans and achieves good performance in the real-world domains we have investigated. We demonstrate the technique on the task of grasping of novel laminar objects and on orienting parts in a tiltable tray.
Issue Date:1993
Description:129 p.
Thesis (Ph.D.)--University of Illinois at Urbana-Champaign, 1993.
Other Identifier(s):(UMI)AAI9314845
Date Available in IDEALS:2014-12-16
Date Deposited:1993

This item appears in the following Collection(s)

Item Statistics