Title:
Failure-Driven Learning as Model-Based Self-Redesign
Failure-Driven Learning as Model-Based Self-Redesign
Author(s)
Stroulia, Eleni
Advisor(s)
Editor(s)
Collections
Supplementary to
Permanent Link
Abstract
Learning is a competence fundamental to intelligence. Intelligent agents who
solve problems in a realistic environment need to learn in order to improve
their performance in terms of the quality of the solutions they produce, the
efficiency of their problem-solving process, and the class of problems they
can solve. Failures in problem solving signify the need and the opportunity to
learn. One way in which an agent may effectively use its failed problem-solving
experiences to learn, is by reflection upon its own problem-solving process.
To that end, the agent needs an explicit model of its own problem-solving
behavior. This work adopts a design stance towards reflective, failure-driven,
learning. This stance gives rise to a specific computational model which is
based on three key ideas: (i) agents can be viewed as abstract devices; (ii)
their problem solving can be understood in terms of structure-behavior-function
(SBF) models; finally, (iii) failure-driven learning can be viewed as a
model-based redesign process, in which the agent uses its comprehension of
its own problem solving to repair itself. When the agent fails, it uses
feedback from the world, and the trace of the failed process, to search
through this model and identify the cause(s) of its failure. Then, it
proceeds to repair its problem solving, in order not to fail again for the
same reason.This theory of reflective learning has been implemented in a
fully operational system, AUTOGNOSTIC. AUTOGNOSTIC is like a "shell" in
that it provides the SBF language for specifying a problem solver, and the
inference mechanism for monitoring this problem-solver's reasoning,
assigning blame when it fails, and repairing it appropriately. Three
different systems have been modeled in AUTOGNOSTIC's SBF language: ROUTER, a
path planning system, KRITIK2, a design system, and an autonomous, reactive
agent implemented in the AuRA architecture. Extensive experiments conducted
with AUTOGNOSTIC demonstrate the generality and the effectiveness of its
learning process.
Sponsor
Date Issued
1995
Extent
1140763 bytes
Resource Type
Text
Resource Subtype
Technical Report