The University of Edinburgh -
Division of Informatics
Forrest Hill & 80 South Bridge

Research Paper #524

Title:Recovering from Plan Failure Using a Layered Architecture
Date: 1991
Presented:Submitted to the 10th UKPlanning SIG, Cambridge, April 1991
Abstract:Planners which are to operate in uncertain environments must be able to recover from failures when they occur. Little work has been done in the past on incorporating recovery strategies into planning systems and deciding which recovery strategy to use when. This paper describes a number of recovery strategies which are useful either generally in planning domains or in multiple agent tasks. It discussed Moore's [3] system, which uses a rigid ordering to choose the next recovery strategy to try, and explains why this technique produces undesirable results for both traditional planning domains and the simulation of human planning. Then it discusses a simple extension to Moore's work [1] which corrects the problem by allowing a recovery strategy to be chosen based on an estimate of the effort needed to complete the task using that strategy, but which is difficult to work with because it "hard-wires" the decisions into the code. A change in approach is required which makes the recovery strategies techniques about which the system makes decisions, rather than outside procedures that the system calls upon as needed. In effect, the recovery strategies, along with normal planning and execution, make up a set of meta-operations which can be planned analogously to domain level plans. The paper describes a system which operates in this way by using a layered planning architecture modelled on Stefik's MOLGEN [5]. By continuing the abstraction of layers upwards, further levels of operators meta-plan the choice of recovery strategy and the choice of the way of choosing a recovery strategy. This system has the advantage that it is easy to both change the way in which recovery strategies are employed and add new recovery strategies, simply by changing the operator definitions at the higher levels. The system is implemented in a task-oriented dialogue domain.

[Search These Pages] [DAI Home Page] [Comment]