MVILAIN@G.BBN.COM (Marc Vilain) (11/11/87)
BBN Science Development Program
AI Seminar Series Lecture
GENERATE, TEST AND DEBUG: A PARADIGM FOR SOLVING
INTERPRETATION AND PLANNING PROBLEMS
Reid Simmons
MIT AI Lab
(REID%OZ.AI.MIT.EDU@XX.LCS.MIT.EDU)
BBN Labs
10 Moulton Street
2nd floor large conference room
10:30 am, Tuesday November 17
We describe the Generate, Test and Debug (GTD) paradigm and its use in
solving interpretation and planning problems, where the task is to
find a sequence of events that could achieve a given goal state from a
given initial state. The GTD paradigm combines associational
reasoning in the generator with causal reasoning in the debugger to
achieve a high degree of efficiency and robustness in the overall
system. The generator constructs an initial hypothesis by finding
local domain-dependent patterns in the goal and initial states and
combining the sequences of events that explain the occurrence of the
patterns. The tester verifies hypotheses and, if the test fails,
supplies the debugger with a causal explanation for the failure. The
debugger uses domain-independent debugging algorithms which suggest
repairs to the hypothesis by analyzing the causal explanation and
models of the domain.
This talk describes how the GTD paradigm works and why its combination
of reasoning techniques enables it to achieve efficient and robust
performance. In particular, we will concentrate on the actions of the
debugger which uses a "transformational" approach to modifying
hypotheses that extends the power of the "refinement" paradigm used by
traditional domain-independent planners. We will also discuss our
models of causality and hypothesis construction and the role those
models play in determining the completeness of our debugging algorithms.
The GTD paradigm has been implemented in a program called GORDIUS. It
has been tested in several domains, including the primary domain of
geologic interpretation, the blocks world, and the Tower of Hanoi
problem.
-------