This paper aims at clarifying the role of problem-solving plans in modeling the process of plan and goal recognition of a speaker. A first step toward the introduction of problem solving plans was made by [Ramshaw1991] with the main goal of characterizing the difference between the activity of plan development and plan execution. The idea is that
in some cases an utterance does not aim at finding out the best way to reach a goal (exploration) but at pursuing an already adopted plan (execution). So, he introduced an independent ``exploration level", able to model the activity of comparing alternative plans, which is entered just in case the speaker is actually trying to decide which plan to adopt.
Some difficulties of Ramshaw's approach were pointed out in [Carberry, Kazi, & Lambert1992]. In particular, the top-down character of his algorithm, which makes difficult to analyze bottom-up dialogues, and the sharp separation of the two phases of exploration and execution, which makes difficult to account for contingent commitments. So, they proposed a ``tripartite" model, which includes domain, discourse, and problem solving goals.
The goal of this paper is to refine the role of problem-solving knowledge, giving it a more prominent role in the activity of goal satisfaction and plan formation. One of the recipes used by Carberry et al. to describe their model (the Domain recipe that specifies how to get a minor in a subject) includes as preconditions ``have-plan (agent, Get-Minor(agent, subj))'' and ``Explored(agent, Get-Minor(agent, subj))''. We claim that such information need not be stated explicitly in the domain actions, since it comes out from the general organization of problem-solving plans. More important, the relation between the three components seems too rigid to account for the flexible relation between the various activities. They introduce three levels (one for each component), since ``discourse actions are performed in order to provide the planning agent with the information necessary to perform a problem-solving action, and problem-solving actions are performed in order to construct a domain plan that can subsequently be executed" ([Carberry, Kazi, & Lambert1992], p. 193). In other words, the problem-solving level plays the role of a link between the domain and the discourse level. We believe that a discourse action (a question, for instance) is performed as the result of a problem-solving activity suggesting that asking is the best way to get some information exactly in the same way as a given decomposition is chosen as the best way to perform a domain action and exactly in the same way as a given politeness form is chosen to perform the speech act. To be clearer, we substitute to the model
Domain level 
PS-level
Discourse level
the model
Goal 
PS-level 
Action
where our model is applied at each of the three levels of the original model:
a domain goal can be reached by decomposing it (via the problem-solving
activity) into subgoals; some of them can be executed directly, others require
additional information, so the ``action" in our model can either be a domain
subaction or can be a discourse action. Of course, nothing says a-priori that
when we engage in a dialogue we are carrying out an exploration activity: if I
say ``Where is the library?" I am probably committed to a plan involving going
to the library. This point is set out clearly in [Ramshaw1991]. But he has
to introduce an explicit marker to keep apart ``exploration" linguistic actions
from ``execution" linguistic actions. It seems that the problem solving level
is able to explain the difference without adding extra information: in the first
case it specifies that a given action (asking X) can be performed for solving
some domain problem, in the other for solving an exploration
problem.