Many text planning systems are being used quite successfully today. Their success, however, has come about as a result of several compromises. Constraining the domain and text types are the most obvious. Related to this, however, are several control issues that have been hidden by the simplified nature of previous systems but are now becoming important as those simplifications are lifted:
`` Most current discourse planners ... rely on customized planning algorithms with procedural semantics for the purpose of solving specific text-planning problems. ... careful analysis of these programs show that there is nothing in their semantics to prevent them from generating incorrect plans, generating plans with redundant steps, or failing to find plans in situations where they exist. To the extent that these planners have been able to avoid these problems, they have done so by severely limiting the expressive power of action descriptions and/or requiring the designer of action descriptions to handcraft each description to fit correctly into the ad hoc semantics of the specific plan for which the action is intended.'' (Young & Moore, 1994)
``With simple state-based representations, complete search strategies will generally be exponential as a function of solution length. With more expressive representations ... determining if solutions to arbitrary problems exist is an undecidable problem. Such disconcerting results have led several researchers to abandon the use of explicit or declarative problem representations. However, it appears that doing so requires that the goals of the agent be within a narrow range that are hard-coded into the problem representation.'' (Tenenberg, 1991).
``Time to impact?'' (Captain J.L. Picard)
The last quote above graphically illustrates what the first two quotes are talking about. When Capt. Picard asks how long his spacecraft has until it is obliterated by alien fire, he needs to know NOW. Furthermore, he needs to know CORRECTLY. Unfortunately, the current generation of text planners are not able to process real-life problems quickly, nor are they guaranteed to process them correctly.
Tenenberg states the obvious problem for all AI applications: basic search strategies have exponential time complexity. Young and Moore point out that most text planning systems currently are neither sound (guaranteed to give correct answers) nor complete (guaranteed to find correct answers). Both citations agree that current approaches sidestep these problems by abandoning declarative knowledge in favor of ad hoc procedures. Young and Moore go on to argue that the proliferation of procedural knowledge leads to unsoundness and incompleteness, and both papers conclude that such an approach can only be successful on a narrow range of limited problems.
Young and Moore, in their paper, go on to introduce the DPOCL text
planning system. The main goal of that research was to ensure
soundness and completeness. In this they no doubt succeeded;
however, no claims were made concerning the
efficiency of their work. Tenenberg, on the other hand, addressed
efficiency in his discussion of abstraction in planning.
We agree with Young and Moore's conclusion that ad hoc procedures
contribute to unsoundness and incompleteness. Declarative knowledge
which clearly marks preconditions and effects must be used, along with
a control mechanism that ensures soundness and completeness. PICARD
can be viewed as an attempt to add efficiency to this type of control
paradigm by applying techniques similar to Tenenberg's abstraction.
The work builds on the
HUNTER-GATHERER analysis system described above. That system
employs constraint satisfaction,
branch-and-bound and solution synthesis techniques to produce near
linear-time processing for knowledge-based semantic analysis in the
Mikrokosmos Machine Translation Project. PICARD enables similar
results for text planning by recasting localized means-end planning instances into
abstractions
connected by usage constraints that allow HUNTER-GATHERER
to process the global problem as a simple constraint satisfaction
problem.