p.275-276 Planning is a familiar cognitive activity. We all have many opportunities to decide how we will behave in future situations... We define planning as the predetermination of a course of action aimed at achieving some goal. It is the first stage of a two-stage problem-solving process. The second stage entails monitoring and guiding the execution of the plan to a successful conclusion. We refer to these two stages as planning and control. We have two main objectives: to characterize the planning process and to propose a theoretical account of it.
[JLJ - alternatively, plans in a complex environment aim for a robust or resilient position where multiple project options are available (and selectable) as further information becomes available, including information not immediately known now in specific detail. Our basic "goal" is to be diagnostically "ready" for whatever happens, by aiming to develop multiple, leveraged projects. Our imagined project scenarios show us successfully maneuvering in the life world to satisfy our needs. Our "plan" might be to refine these a project-based systems as the future world emerges/develops/changes, satisfying our needs as they emerge, turning our concerns into a portfolio of projects, which we then resourcefully and skillfully execute, using either scripts that "work" or improvisations which "likely will work", or even just "showing up" at places and times where events will likely further develop to our benefit.]
p.276 Sacerdoti's (1975) work [JLJ - A structure for plans and behavior. Technical Note 109, Standford Research Institute, Menlo Park, California, August, 1975] is probably the best-known previous research on planning. His computer program, NOAH, implements a successive refinement approach to planning. NOAH formulates problems in terms of high-level goals that specify sequences of actions (for example, the monkey should get the bananas and then eat them). NOAH expands each constituent subgoal into additional subgoals, maintaining any indeterminate sequential orderings as long as possible. In this manner, NOAH eventually generates correct plans specifying sequences of elementary actions. When executed, these actions transform initial conditions into a series of intermediate conditions, culminating in the goal state. (See also: Ernst & Newell, 1969; Fahlman, 1974; Fikes, 1977; Fikes & Nilsson, 1971; Sacerdoti, 1974; Sussman, 1973).
While not incompatible with successive-refinement models, our view of planning is somewhat different. We share the assumption that planning processes operate in a two-dimensional planning space defined on time and abstraction dimensions. However, we assume that people's planning activity is largely opportunistic. That is, at each point in the process, the planner's current decisions and observations suggest various opportunities for plan development. The planner's subsequent decisions follow up on selected opportunities... In general, the assumption that people plan opportunistically implies that interim decisions can lead to subsequent decisions at arbitrary points in the planning space. Thus, a decision at a given level of abstraction, specifying an action to be taken at a particular point in time, may influence subsequent decisions at higher or lower levels of abstraction, specifying actions to be taken at earlier or later points in time.
p.285-286 The proposed model generalizes the theoretical architecture developed by Reddy and his associates... for the Hearsay-II speech-understanding system... Others have since applied it to... protein-crystallographic analysis (Nii & Feigenbaum, 1977)... The proposed model is, to our knowledge, the first attempt to adapt the Hearsay-II architecture to a "generation" problem.
p.286 As mentioned above, independent cognitive specialists generate decisions during the planning process. The model operationalizes specialists as condition-action rules.
The condition component describes the circumstances under which the specialist can contribute to the plan... The action component defines the specialist's behavior... Thus specialists generalize the symbol-manipulation capabilities of production rules (Newell & Simon, 1972) to more complex, pattern-directed activity
p.291 Under the control of the executive, the planning process proceeds through a series of "cycles" during which various specialists execute their actions. At the beginning of each cycle, some number of specialists have been invoked - that is, their conditions have been satisfied. The executive selects one of the invoked specialists to execute its action - that is, to generate a new decision and record it on the blackboard. The new decision invokes additional specialists and the next cycle begins. This process ordinarily continues until: (a) the planner has integrated mutually consistent decisions into a complete plan; and (b) the planner has decided that the existing plan satisfies important evaluation criteria.
p.302 The opportunistic model seems, at first glance, fairly complex. It postulates five different conceptual "planes" of decisions and several levels of abstraction within each of those planes. It postulates numerous planning specialists whose simultaneous efforts to participate in the planning process require the supervision of a fairly sophisticated executive. Although a number of complex models have proved fruitful in the last few years (Cf., Anderson, 1976; Anderson & Bower, 1973; Rumelhart, Lindsay, & Norman, 1972; Winograd, 1972), most of us still adhere to the law of parsimony, preferring simpler models to complex models.
In fact, the proposed model is computationally quite simple. It postulates a uniform decision mechanism, the specialist, to perform all of the varied decision-making functions planners perform.
p.303 Most of the apparent complexity in the model derives from the details of the blackboard structure. However, the blackboard partitions provide another important computational efficiency. Each specialist gets invoked whenever a new decision on the blackboard satisfies its condition... The blackboard partitions reduce the amount of computation required by permitting each specialist to restrict its "attention" to only those new decisions that occur at particular levels.
The blackboard structure also permits the model to capture an important psychological feature - interruptibility. People have the power to interrupt their own cognitive processing at arbitrary points. After performing some more or less related processing, they may or may not continue the interrupted task. This interruptibility appears throughout our protocols and we believe it is a salient feature of cognitive processing in general.
p.303 The diverse observations people make while planning often guide subsequent planning.
p.306 planning in tasks fraught with complexity and uncertainty might benefit from less of the discipline imposed by a top-down process. In such complex tasks, general, a priori solutions or problem-solving methods may not exist or may be computationally intractable. Even if some general approach were available, opportunistic planning would free the planner of the burden of maintaining a structurally integrated plan at each decision point. Instead, the planner could formulate and pursue promising partial plans as opportunity suggested.
More importantly, a multi-directional process might produce better plans. It certainly permits more varied plans than a strictly top-down process does... The bottom-up component in multi-directional processing provides a potentially important source of innovation in planning. Low-level decisions and related observations can inspire novel higher-level plans... Feitelson and Stefik (1977) observed that their expert geneticist deliberately exploited the potential for innovation in bottom-up processing:
Thus, not only is the planning process largely event driven but sometimes steps are taken somewhat outside the plan of the experiment to make a possibly interesting observation. This kind of behavior reflects the convenience of making certain interesting observations while the equipment is set up. Often this is done to verify the successful completion of an experimental step, but sometimes the observations seem to correspond more to fishing for interesting possibilities. (p. 31)
p.307 A study by Hayes-Roth (1979) provides more direct evidence for the influence of problem characteristics. She successfully induced alternative planning approaches by manipulating the amount of time available for plan execution. For problems that imposed severe time constraints, most subjects adopted a top-down approach. For problems that imposed minimal time constraints, most subjects adopted a bottom-up approach.
p.307 Feitelson's and Stefik's (1977) study of the experiment-planning of an expert molecular geneticist provides a nice illustration:
The experiments described here reflect a combination of goal driven behavior and event driven behavior... If there were no goals, behavior might seem very erratic and follow no general course. If there is no event driven component to the planning process, then the experimental procedure must admit no feedback or changes of plans as a result of observations. Thus, no advantage will be made of fortunate observations. What is being suggested here is that the planning in this experiment involved far more exploitation of events and changes of plan according to the events than the authors had anticipated. (p. 30)
p.308 CONCLUSIONS The opportunistic model draws on earlier theoretical work in cognitive psychology and artificial intelligence. It incorporates the strongest points of these models with its own assumptions regarding multi-directionality, opportunism, and incrementation in a heterarchical plan structure. We believe that the model is flexible enough to handle the complexity and variability of people's planning behavior. Yet, it is vulnerable to data. We hope the opportunistic model will provide a useful framework for future investigations of the planning process. [JLJ - interesting conclusion, we "hope" our research results will be useful, much as we "hope" that the opportunistic plans we develop will be useful.]
