<?xml version="1.0" standalone="yes"?>
<Paper uid="H93-1033">
  <Title>Generic Plan Recognition for Dialogue Systems</Title>
  <Section position="6" start_page="175" end_page="175" type="concl">
    <SectionTitle>
5. Discussion
</SectionTitle>
    <Paragraph position="0"> Graph-based approaches to representing plans date back to the very beginnings of work on automated planning, from Sacerdoti's procedural nets \[6\] to SIPE's representation of plans \[7\]. Often these representations reflected a combination of temporal information and knowledge about the plan.</Paragraph>
    <Paragraph position="1"> In our view, the temporal reasoning is provided by the underlying knowledge representation and the plan graph represents an argument that a certain course of action under conditions will achieve certain explicit goals. The earlier systems' inability to separate the plan representations from their use as data structures in planners made it difficult to predict and explain their behaviour. The plan graph formalism achieves such a separation, but the price we pay is the inability to use directly the efficient algorithms developed previously. Some of the results from the planning community on efficient algorithms can be adapted to the temporally explicit logic of events (c.f., \[2\]). We are developing a theory of plan graphs that will provide a formal basis for many of the heuristic procedures developed previously.</Paragraph>
    <Paragraph position="2"> With respect to plan recognition, Kautz's work \[8, 9\] provides a formal basis for plan recognition but only dealt with observed events fitting into a hierarchy of event types. Pollack \[10\] uses a formalism similar to our underlying temporal logic, but includes representations of belief and intention that are not the focus of this paper. We believe that there is a structure to plans independent of the intentions of agents, and that plan graphs seen as arguments provide the proper perspective for reasoning about them at that level.</Paragraph>
    <Paragraph position="3"> Carberry \[11\] describes a model for incremental plan inference in task-related information-seeking dialogues. It uses a &amp;quot;context model&amp;quot; consisting of a tree of goals, with associated plans. Since we see the overall structure of a plan as an argument, there is no such separation in our approach, although we do treat goals specially as described previously. Her &amp;quot;current focused&amp;quot; goal and plan are analogous to our :purpose mechanism and to the techniques used by the language and discourse modules for determining focus. The system also uses breadth-first search with &amp;quot;focusing heuristics,&amp;quot; several of which correspond to our heuristics described previously.</Paragraph>
    <Paragraph position="4"> However, the approach lacks a formal description that we believe can be provided by the plan graph formalism.</Paragraph>
    <Paragraph position="5"> Several recent approaches to plan recognition \[12, 13\] rely on the use of a powerful terminological reasoner to place event types in a virtual lattice. This has the advantage that subsumption relationships (corresponding to our unification procedure) can be automatically and incrementally computed.</Paragraph>
    <Paragraph position="6"> Existing terminological reasoners, however, typically either do not allow complex objects (roles) and equality, or draw only the conclusions about subsumption that are deductively (necessarily) entailed. Neither do they compute the assumptions that would unify facts.</Paragraph>
    <Paragraph position="7"> No existing system is as ambitious as the TRAINS domain plan reasoner in providing services required to support dialogue, from representing complex, partial and incorrect plans to providing incremental and interleaved planning and plan recognition. We are currently completing a new implementation of the procedures based on this paper. It is part of our current research to apply work on argument systems directly to justifying these plan graph algorithms in terms of a formal theory of plan graphs.</Paragraph>
  </Section>
class="xml-element"></Paper>