The method of representing uncertainty in CNLP and PLINTH has important implications for how knowledge goals are handled in their plans.
The acquisition of information is a planning task like any other [Pryor and Collins1991, Pryor and Collins1992, Pryor 1994]. In general, the sequence of actions required to achieve a given knowledge goal may be arbitrarily complex. For example, an action to observe a tossed coin might require that the observer is in the appropriate location; in other cases, there might be several different possible methods of information gathering, some involving perception, some involving reasoning, and some a combination. A contingency planner, some of whose plans will necessarily involve the achievement of knowledge goals, must therefore be able to plan fully generally for information gathering.
The confusion between the source of uncertainty and the observation of uncertain results limits the ways in which knowledge goals can be achieved in CNLP and PLINTH: they must be achieved through the special observation actions that specify the uncertain outcomes. This is a result of their representation in terms of the planner's world model, which means that they do not represent the effects of actions (except to flag them as unknown) until the planner has observed them (or otherwise incorporated them into its world model). In their discussion of this issue Goldman and Boddy [1994b] explicitly exclude knowledge goals from consideration. As they point out, planning under uncertainty requires that a distinction be made between the actual state of the world and the planner's knowledge of it. In order to plan effectively for knowledge goals, both must be represented. This is done in Cassandra by separating the representation of uncertainty from the representation of information-gathering. If an effect results deterministically from an action, Cassandra reasons that there is no need to observe it, and it forms part of the world model. An uncertain effect, on the other hand, is not incorporated unconditionally into Cassandra's world model; it is noted as being possibly true, and (if necessary) Cassandra sets up a subgoal to determine whether it is indeed true.
SENSp, which uses the UWL representation for goals and actions,
has three different kinds of precondition that can be used to
represent information goals either alone or in combination [Etzioni et al. 1992]. As well as
satisfy preconditions, which may be achieved through
actions or through observation, UWL has hands-off
preconditions indicating that the value of propositions must not be
changed in order to achieve the subgoal, and find-out
preconditions. The latter are in some ways similar to preconditions
for know-if propositions in Cassandra. A precondition
such as (find-out (P . v)) tells the planner to ascertain
whether or not P has truth value v. Under
certain circumstances this type of precondition may be achieved by an
action that changes the value of P. Knowledge goals may
thus be represented by find-out preconditions or
satisfy preconditions (often used in conjunction with
hands-off preconditions). Etzioni et al. argue that
knowledge goals should only be achieved through actions that change
the value of the proposition in question when that change is required
for another purpose in the plan. We believe that this is an
unnecessary limitation, and that in some circumstances enforcement
actions may be the best way of achieving knowledge goals.