[ CogSci Summaries home | UP | email ]

B. Hayes-Roth and F. Hayes-Roth, A Cognitive Model of Planning. International Joint Conference on Artificial Intelligence, 1979.

Author of the summary: Jim Davies, 1999, jim@jimdavies.org

Cite this paper for:

note: page numbers are from the version in "Readings in Cognitive Science."

Previous models of planning like Sacerdoti (1975) tend to be successive-refinement models. This model shares the view that the planning process happens on a 2d space of time and abstraction.

Presented is a protocol of planning. In it the subject changes levels of abstraction throughout. At one point the subject notices that many errands in the task occur in same area, so the subject makes a higher level plan to go to that area and do all of those errands at the same time. The first thing the subject does in place all items in the plan into 2 categories, primary and secondary, according to importance. The subject does not plan strictly forward in time either. The subject sometimes works on the details of subplans. The subject also uses mental simulation to get information about how long things should take, etc. (p500).

This paper presents an opportunistic model of planning. Planning is done by a bunch of specialits (akin to demons in Selfridge's Pandemonium model (1959)). All specialists record decisions on a shared memory structure called a blackboard. Specialists make decisions based on heuristics and what is on the blackboard.

The blackboard is broken up into planes, each with different conceptual information (e.g. explicitly planned activities, decisions about data.) Each plane has multiple levels of abstraction. Most specialists deal with only a few levels of a few planes.

Specialists are like production rules. They have complex preconditions and the postconditions always result in a new or changed decision. The blackboard contains 5 conceptual planes: plan, plan abstractions, knowledge base, executive and meta plan.

plan plane: specific actions the planner will carry out.
Abstraction levels:

  1. outcomes: what the planner intends to accomplish with the completed plan.
  2. designs: general behavioral approach like the general order according to area.
  3. procedures: specific sequences of gross actions.
  4. operations: order of minute actions.

plan abstractions plane: desired attributes of potential plan decisions. (e.g. "go to the nearest errand next")
Abstraction levels: (each cooresponds to the levels of the plan plane.)

  1. intention: what kinds of things go in the outcomes level. (e.g. the primary things)
  2. scheme: what kind of design? (e.g. breaking the map up into areas)
  3. strategy: like "go to the closest errand next"
  4. tactic: like "search for a shortcut between these two errands"

knowledge base plane: observations and computations about the world. (e.g. "the florist is closest to the health club.") (each cooresponds to the levels of the plan abstractions plane.)
Abstraction levels:

  1. errand: relative importance of events
  2. layout: convienient spatial clustering
  3. neighbor: two errands are near one another
  4. route: detection of a shortcut

executive plane: decisions about the allocation of cognitive resources. (which part of the plan will we develop now? Which specialists will do it?)
Abstraction levels:

  1. priority: which part of the blackboard to focus on?
  2. focus: when should I make what decisions?
  3. schedule: resolution of conflicts among specialists

meta plan plane: the planner's understanding of the problem, how to evaluate the plan, etc.
Abstraction levels:

  1. problem definition: goals, resources, actions, constraints
  2. problem solving model: how will the problem be represented? (e.g. as the traveling salesman problem, divide and conquer)
  3. policies: global constraints and desirable features of the developing plan.
  4. evaluation criteria

Control (p504)

It happens in cycles. The executive chooses which matching specialist will make a decision. The decision made results in the system being in a different state, and the cycle repeats. This happens until the plan in complete or failed.


The protocol described earlier was modeled. The model had about 40 specialists. (p507) The plans produced by the subject and the model are similar. The main difference in that the model's plan was more realistic with regard to how many things one could get done. The model dropped more secondary items. The model made many decisions made in the protocol. (p509)

There were differences, though. Part of the reason for this is that the model makes random choices among specialists when many are of equal value. Also. the list of specialists is incomplete.


The model seems complex, but the different planes can be looked at as simply partitioning the knowledge. In facs there in only one decision making mechanism. The way it is set up also allows modeling of interruptibilty, or changing trains of thought.

People do not work completely top-down. They make planning decisions at many levels and do not worry about consistency all the time. But both ideas have merit and are good for modeling different things. For example, people seem to plan dinners in a top-down manner.

Summary author's notes:

Back to the Cognitive Science Summaries homepage
Cognitive Science Summaries Webmaster:
JimDavies ( jim@jimdavies.org )
Last modified: Sun May 16 14:01:55 EDT 1999