We discuss how resources affect the decision-making process. We elaborate a general definition of a resource as something that satisfies the enabling condition for a plan. We develop a taxonomy of fifteen types of resources and eight separate dimensions for classifying resources. In this context, we will discuss cognitive resources, such as memory and attention. Our model of goals leads us to a cognitive analysis of emotions and moods.
Goals, plans, and resources are closely associated. This basic triad is depicted in Figure 4.1. The following associations hold among goals, plans, and resources.
Goals give rise to plans. That is, plans provide a means for achieving goals. The goal of hunger may lead to the plan of going to a restaurant.
Plans give rise to goals. Plans often involve the instantiation of sub-goals. The plan of going to a restaurant may lead to the sub-goal of finding a taxi.
Plans require the use of resources. The execution of a plan may involve the use of a resource as a precondition. To take a cab, you need to have enough cash.
Resources give rise to plans. There may be a standard plan for the acquisition of a resource. If you need cash to pay for the cab, you may go to an automatic teller machine.
Resources give rise to goals. Unexpected resources can lead to opportunities — unanticipated goals. For example, while going to the automatic teller, you may notice an interesting restaurant and decide to abandon the goals of getting cash, taking a cab, and going to the other restaurant. The new restaurant is an opportunistic resource which allows goal substitution.
We may use knowledge about the associations among goals, plans, and resources, as well as knowledge about the process of using resources to make inferences about goals.
In a goal tree or similar hierarchical representation, goals are ranked according to importance. The more important a goal, the higher it would appear in the goal tree. An agent’s action might be explained as an attempt either to achieve or preserve some positive goal state or to block some negative goal state. Furthermore, when given a choice between multiple actions, an agent will usually choose that action which is in service of the most important or highest goal or collection of goals.
Given the interdependence between goals and actions, one may often infer one from the other. Given knowledge of an agent’s goals and the choices of actions, one might well predict what action the agent will take, as in the following example.
Example 4.1 John was thirsty. He opened the refrigerator door.
We infer that John will get something cold to drink.
At the same time, by observing what choices an agent makes, one may then infer what goals the agent was trying to achieve.
Example 4.2 John opened the refrigerator door. He took out the milk and got an empty glass.
We infer that John is thirsty.
The notion of the ranking or importance of goals can be given more explicitly by relying on two basic principles for the achievement of goals:
Principle 1 Resources are required to achieve goals.
Principle 2 Agents wish to maximize their achievement of goals.
These principles are similar to Wilensky’s meta-themes [Wilensky, 1981]. A natural consequence of these two principles is that agents wish to maximize their resources. The concept of a resource can be more explicitly defined.
Definition 4.1 A resource is something that satisfies an enabling condition of one or more plans.
We view resources quite generally. Any plan requires certain resources for its achievement. To succeed, an agent must possess these resources. He must have the wherewithal — the necessary means. For example, two common resources are time and money. Many (if not all) goals require some expenditure of time. The absence of resources can prevent the achievement of goals:
Example 4.3 John wanted to go to the movies, but he was broke.
Example 4.4 John wanted to go to the movies, but he had to work late.
The fact that resources have to be expended in achieving goals clearly constrains the choices that a planner must make.
It is also possible in many cases to measure the resources spent in the achievement of a goal. Given the corollary of maximizing resources, we may derive our original principle of importance for inferring goals.
Principle of Importance. The importance of a goal is proportional to the resources that the agent is willing to expend in pursuit of that goal, and vice versa.
Just as the very ranking of goals is subjective, so is the value of resources, as shown in the next examples.
Example 4.5 John spent $100,000 (his life savings) running for governor.
Example 4.6 Nelson spent $10,000,000 (his annual income) running for governor.
These examples demonstrate that resources should not be measured on an absolute scale, but rather viewed subjectively relative to the agent concerned.
We should also point out that the calculation of importance from resources must take into account resources already allocated to a given goal. For example, an agent may expend lots of time or money in pursuit of a goal while still not achieving the intended goal. The importance of that goal at that time is not measured merely by the incremental cost of achieving the goal, but rather should take into account the cumulative resources expended.
Example 4.7 John had worked his way through college and law school. To get enough money to take the bar review course, he would have to sell his car.
The decision for John is not simply whether passing the bar exam is worth losing his car. If John does not pass the exam, he has in effect lost the time, money and effort spent on his legal education. John has invested in his education, which has increased its importance. John must protect that investment.
Principle of Investment. The importance of an active goal is proportional to the resources that the agent has already expended in pursuit of that goal.
We use the term active to indicate that this principle should not apply to latent or pending goals. Still, there is a sense in which we clearly believe current goals to be more important than goals that are not being pursued.
However, there are times when one must abandon an investment. The sayings “do not throw good money after bad” or “cut your losses” refer to this situation. Just because a goal is important does not mean that you will achieve it. A less important goal will be abandoned more readily than an important goal. Thus, it will be harder to abandon a high investment goal, than a goal of low investment.
The principle of investment is descriptive, not prescriptive. Economists argue that true investment decisions should consider only future costs and returns. Past expenditures should be considered “sunk costs” and not affect the decision. We suggest, however, that the principle of investment is a psychological reality which influences decision making. This model is consistent with Prospect Theory of Tversky and Kahneman (1979) which is discussed in chapter 8. [Note: economic decision theory stipulates that an agent should ignore Sunk Costs. However, Tversky and Kahnemann, as always, project psychological evidence and justification.]
Given the specific definition of a resource as something that satisfies the enabling conditions of other goals, there are numerous things that would qualify as resources. Here are examples.
Example 4.8 John was often late for work.
We wish to distinguish three different aspects of time as a planning resource. First, there is time as an amount — the duration of time required for completion of a task. For example, it might take John 15 minutes to drive to work, 20 minutes to take the bus, or 45 minutes to walk.
Second, there is the required completion time — a deadline for finishing a task. This resource reflects the urgency of a task. For example, we might assume that John is supposed to be at work at 9 am. Thus, if he drives, he can leave home at 8:45.
Finally, there is the need to synchronize actions — the timing of an action. For example, if John plans to take the bus and leaves home at 8:40, he may still be late if the bus leaves on the half hour.
By comparing time to other resources, we do not mean to suggest that reasoning about time is simple. AI researchers have demonstrated many subtle and difficult problems associated with temporal reasoning [Allen, 1984, Dean, 1985, Shoham, 1986, Hanks, 1990]. On the contrary, we wish to suggest that reasoning about other resources, such as the ones listed below, will also be complex.
We also want to emphasize that time exhibits many features found in resources in general. Time can satisfy enabling conditions on plans.
Example 4.9 John had enough time to eat lunch before his meeting.
Time limitations can lead to problems that may result in changes in plans.
Example 4.10 John did not have enough time to eat at Chez Che, so he got a burger instead.
A surplus of time can provide a new opportunistic goal.
Example 4.11 John’s meeting was canceled, so he decided to take a walk.
Time is no doubt different from other types of resources in many ways. However, we feel justified in treating time as a resource just the same.
Example 4.12 John bought a new car.
Money, like time, is of no intrinsic value. Money’s value lies in the fact that it can be exchanged for other resources that will satisfy goals. Furthermore, since most goals can be translated into a cost in terms of money, money can provide a convenient means for comparing the value or importance of otherwise disparate goals.
As with time, we recognize that money has many complex and difficult aspects. The fields of economics, accounting, and finance examine many of the subtle and intricate features of money.
Example 4.13 John stopped the car when he saw the man with the white cane.
Sight, speech, hearing, touch, and the use of limbs are all resources that people rely on for a vast number of actions. The lack of one of these abilities is an infirmity which restricts the plans or goals that one may pursue. Someone who is afflicted may resort to substitutes, such as hearing aids or prosthesis. Some physical abilities, such as strength, have a quantitative measure, e.g., “John can lift 300 pounds.”
Example 4.14 John needed a ride to work since he did not know how to drive.
Certain plans require skills. Possession of a skill is a resource needed for pursuing certain plans. Physical skills include typing, handwriting, playing the piano, and sports. Mental skills would include reading, math, and language fluency.
Example 4.15 John needed a ride to work since his license had been revoked.
Society often requires an explicit statement of qualifications or franchise. Suffrage, citizenship, and professional degrees are examples of this type of resource. More generally, an agent may require permission from an authority to perform some action. A parent may forbid a child to watch television.
Example 4.16 John did not need a ride to work since he lived only two blocks from the office.
Many plans require that an agent be in a specific physical location. Therefore, once the agent is in that location, he has satisfied an enabling condition for the plan. In the present example, we can see that John has solved his commuting problem by moving closer to work. The phrase “being in the right place at the right time” captures this notion of location (and time) as a resource. Many plans are described by the initial change in location, e.g., “going to the movies” or “going to a restaurant.”
Example 4.17 John could not find a parking place near the office.
Being in the right place is not always sufficient. Many plans require a certain physical volume to be occupied by people or objects. Consider an efficiency apartment versus a three bedroom house, or a motorcycle versus a station wagon. It is easier to park a motorcycle, but you can carry more in a station wagon. Physical space and volume can be considerations in many types of plans.
Example 4.18 John’s brother gave him a ride to work.
People can help other people. Someone can act on your behalf as your agent. In chapter 5 we will discuss how relationships can be viewed as resources.
Example 4.19 John was too sick to go to work.
Life is the ultimate individual resource. You do not leave home without it. Almost all goals that a person pursues requires a pulse. Different levels of illness will limit plans accordingly.
Example 4.20 John needed a ride to work while his car was in the shop.
We have many tools for achieving disparate goals. Some of the tools give rise to goals of their own. Complex tools have a teleological character. We speak of a car needing gas and oil. We understand that the common design and function of an automobile require us to service its needs. If we viewed our car merely as an immobile sculpture, we would have fewer things to take care of.
Example 4.21 John got lost on his way to his new job.
Knowledge is different from a mental skill. In this example, John knows how to drive and how to follow directions. He simply does not have accurate directions. He is missing the correct plan.
Example 4.22 John grew some vegetables in his back yard.
This is the common notion of resources found in nature. Natural resources would include minerals, lumber, food, coal, oil, clean air, clean water, harbors, and fertile soil.
Example 4.23 John called the mayor to get the potholes fixed.
Position or rank or office provides certain authority. The mayor can order the city street department to fix the potholes. Similarly, if the mayor is indebted to John for any reason, John can effectively order the mayor. That is power. In the following chapter on interpersonal relationships, we discuss the process of persuasion, one component of which is the power relationship between agents.
We do not suggest that the preceding list is exhaustive. It is merely illustrative. The reason for viewing this disparate collection of conditions all as resources is to provide a uniform way of dealing with them to infer the importance of goals.
The types of resources described in the preceding section suggest the following set of features for depicting resources. Table 4.1 categorizes the preceding types of resources along these dimensions.
Perishable. Some resources are persistent and durable, such as knowledge, abilities, credentials, and relationships (especially familial). Others are more perishable or transient, such as location, and time. The degree of transience depends on the time scale. For example, viewed in the large, life is clearly transient, which implies that knowledge, abilities, etc. are likewise transient. Thus, resource persistence has to be defined relative to a time period. Note that we normally view opportunities as transient.
Table 4.1 addresses the question “Does this type of resource degrade over a period of 1 day, or over a period of 10 years?”
Expendable. Some resources disappear or degrade with use. Money is a prime example. People employ the language of money to refer to expendable resources, using words like pay, spend, waste, and afford. Note that we distinguish expendable resources from perishable resources. Expendable resources degrade with use. Perishable resources degrade with time. The two often go hand in hand, but not always. Money is expendable, but it is not perishable, in our specific sense, except as its value may decrease due to inflation. On the other hand, a credential can be perishable, but not expendable. Driving a car does not slowly diminish the validity of your driver’s permit. However, the permit itself will eventually expire and need to be renewed.
Table 4.1 addresses the question “Does this type of resource degrade with use?”
Critical vs. Fungible. Some resources, such as money, are interchangeable. It often makes little difference if an agent uses one form of currency or another to pay a bill. Other resources, such as time and location, are less flexible. You cannot buy groceries after the store is closed or if you have gone to the wrong store to start with. A critical resource is one that is not interchangeable. Thus, a critical or unique resource can be of much greater importance than a fungible resource. Critical resources might include a passport, your health, or control of the Panama Canal. A planner has to be able to develop alternative plans in the event of failures in plan execution. In that regard, plans that do not require critical resources are generally more reliable, and therefore preferable to plans which depend on unique resources.
Table 4.1 addresses the question “Can this type of resource usually be interchanged with another type of resource?”
Cost and Access. Often, resources are proportional to importance: The highest goals require the greatest resources. Easy tasks require few resources (low cost), and have fairly certain outcomes. Hard tasks require many resources (high cost), and may not succeed. Actually, the difference between easy and hard is not simply few versus many resources. It is accessible/inexpensive versus inaccessible/expensive resources. The increased uncertainty of hard tasks means a greater potential resource expense — possibly infinite if a plan always fails and the goal is not abandoned.
Table 4.1 addresses the question “Is this type of resource usually accessible?”
Acquisition and Renewal. Associated with a resource are ways in which to acquire that resource. Money can be gotten in many ways. Time is different. We cannot turn back the clock. Time comes and goes no matter what we do — although we can effectively create time by getting someone else to do things for us. There are stereotypical plans for acquiring certain resources, such as getting a college degree, becoming mayor, learning to play the piano, buying a car, or getting a date.
Table 4.1 addresses the question “Is this type of resource usually renewable?”
Interleaving. Some resources can be used in more than one plan at one time. The resource can be interleaved. Time and location are two examples. You can do many things at once at a given location. Money, in its normal use, is a resource that cannot be interleaved or shared among plans.
Table 4.1 addresses the question “May this type of resource be interleaved among multiple plans?”
Plans and Contexts. Each resource has indices to useful plans that require that resource. Plans would also have contexts that would be the set of conditions under which a plan would normally achieve its respective goal. Similarly, resources have a context with respect to particular plans. For example, US currency may be used for making purchases, except in countries where it is not accepted. A ubiquitous resource like time would not index all plans; that would be useless. However, there should be contexts or subcategories of time that index certain plans. For example, things to do on a weekend or while waiting for the dentist.
Table 4.1 addresses the question “Is this type of resource used in all contexts or merely some?”
Proprietary and Transferable. Most resources belong to the individual in some sense. This is clearly true for tangible resources like money and objects, but it is also true of skills, knowledge, power, and time. People tend to speak possessively about resources, e.g., “my time” or “my thoughts.” As we consider interpersonal relations, we shall notice that some resources, such as money, objects, and knowledge, can be transferred to other people, whereas other resources cannot be given away, such as location, abilities, and skills.
Table 4.1 addresses the question “May this type of resource be transferred to other agents?”
In the preceding chapter we discussed goal comparison as a central part of decision making. When faced with multiple goals, an agent has to decide which goals to pursue. The agent must compare the goals. We argued that goal importance is a critical part of this process. Moreover, the use of scales for comparing or measuring psychological phenomena is mirrored in large part in the goal hierarchy. Scales emerge from the ranking and comparison of goals.
Resources reflect similar phenomena. We find it natural to compare resources and make judgments and decisions based on these comparisons.
Consider a common functional object: an automobile. What makes one car better than another car? The comparison of one car to another will depend on the goals of the agent. Here is a set of possible goals that a prospective car buyer might have.
Speed. The agent likes to drive fast.
Price. The agent does not wish to spend too much money.
Financing. The agent does not care how much it costs, as long as he or she does not have to pay much up front.
Safety. The agent wants a model with a good safety record.
Reliable. The agent wants a car that will not break down on a long trip.
Cost of Ownership. The agent wants a car that is inexpensive to maintain.
Legroom. The agent is six and a half feet tall.
Cargo Space. The agent needs room for hauling building materials.
Image. The agent wants a car like the one his neighbor drives.
Dealer. The agent wants to buy a car from his brother, the car salesman, not to get a better deal, but to help his brother.
You cannot say that car X is better than car Y without having some set of goals that the car is meant to satisfy. For many resources, though, we associate normative sets of goals. We understand why it is good to have money and that more money is better than less. However, we can imagine situations in which it is bad to have lots of money. For example, the rich are more likely targets for kidnapping and extortion in general.
The words good and bad, together with their comparative forms, are often used to express implicitly such normative knowledge. For physical abilities, we speak of good vision, or bad hearing. John may be a better tennis player than Mary, while Mary is a better driver than John. These type of judgments implicitly refer to sets of goals that these resources either do or do not satisfy. Furthermore, these comparisons reflect a ranking of these goals.
Example 4.24 John could parallel park, but was never able to drive on the interstate without causing an accident.
Example 4.25 Mary always had trouble parking, but had never been involved in an accident.
Here, both John and Mary have imperfect driving abilities, but we can conclude that Mary is a better driver since her deficiency is minor compared to John’s. We base this judgment on an implicit ranking of goals: Causing traffic accidents is worse than having parking problems.
In the previous chapter we discussed the use of episodic knowledge in providing us with normative goal knowledge for various categories of people. That is, we can infer certain of a person’s goals by knowing something about a person’s age or job or background. We might also be able to infer something about a person by knowing the person’s goals.
This type of stereotypical information is also known for resources. That is, we make assumptions about what resources may be available to a given person based on limited knowledge of that person. For example, for the chronological categories provided for normative goals in the preceding chapter, we offer the following normative resources.
Infant. Lots of time, limited physical abilities, proximity to parents.
Child. Ability to walk and talk, basic social skills.
Teenager. Ability to play games and sports, limited money, excess energy, basic academic skills such as reading and math.
College student. Advanced academic skills, such as ability to write papers and read a foreign language, advanced social skills, credentials, such as a driver’s license or voter’s registration, more functional objects such as a stereo.
Adult. Job skills, more money, domestic skills such as shopping and cooking, financial skills such as balancing a checkbook, filing a tax return or applying for a loan, more functional objects such as a car.
Married adult. More money, less time, a house, a spouse, household functional objects such as appliances.
Parent. Less time, less space in house.
Grandparent. More time, more space in house, poorer health.
These examples are meant to be suggestive, not absolute. As with default goals, other dimensions, such as occupation or education or culture, could be used to delineate default resource knowledge.
This type of normative knowledge provides us with sets of expectations. For example, we expect a bartender to make more money than a college student, but considerably less than a corporate lawyer. If we read about a “rich bartender,” we are puzzled. This is an anomaly. This is an expectation failure. It should trigger the process of explanation [Schank, 1986].
We have discussed functional objects as resources. However, if we view these objects as reflecting the goals of their design, as we did in the preceding chapter, we can see that functional objects may themselves possess sets of resources. We return to our examples of cars and computers. They each have a set of default resources.
Car. Gas, oil, rust-free body, windshield, tires, radio, windshield wipers, heater, defroster, exhaust system, fan belts, brake fluid, spark plugs, headlights, rear-view mirrors, battery, etc.
Computer. Main memory, disk drives, tape drives, printers, console, operating system, power supply, network hardware, etc.
The relative importance of the component resources will again depend upon the importance of the related goals. Thus, you can drive your car if the heater or radio is broken, but not if the tires are flat.
Where do goals come from? How do people develop new goals?
We suggest two primary types of goal development: top-down and bottom-up. In top-down goal development, high-level goals give rise to sub-goals which are instrumental to the achievement of the higher goals. For example, the high-level goal be-elected-mayor generates sub-goals such as raise-money, give-campaign-speeches, and give-press-conferences.
The second type of goal development is bottom-up, or opportunistic. When an agent recognizes the presence of a resource, the agent may then adopt a plan that requires that resource. That plan may not be instrumental to any current high-level goal. Rather, the agent is behaving in an opportunistic fashion. The resource represents an opportunity, not a planned event. For example, a political candidate might receive an unsolicited endorsement which can then be used for generating further publicity and support.
We may term these two types of planning goal-based and resource-based. The top-down approach to planning is goal-based, whereas the opportunistic, bottom-up approach is resource-based.
One salient difference between the two approaches relates to the likelihood of success. Everything else being equal, bottom-up planning should have a higher rate of success than top-down planning. In resource-based planning, the agent is initiating a plan based on the knowledge that a resource required for achieving the plan is present. That is, the plan is already partially satisfied. In goal-based planning, there is no such guarantee. It is easy to imagine situations in which an agent wants to achieve some goal for which major preconditions are not met.
In discussing the relationship between goals and resources, we note the following general process for the use of resources in plans.
. Acquisition. The agent gets control or possession of the resource.
Example 4.26 John unlocked his car.
. Preparation. The agent performs preliminary actions to ready the resource for use.
Example 4.27 John started the car.
. Action. The agent performs the intended actions with the resource.
Example 4.28 John drove to the restaurant.
. Restoration. The agent returns the resource to a state ready for future use.
Example 4.29 John parked the car, turned off the engine, and locked the door.
In goal-based or top-down planning, the agent must plan for the acquisition of the resources. That is, the agent must do whatever is required to obtain the resource, be it time, money, or a functional object, such as an automobile. There is always some chance of failure associated with the acquisition step.
In resource-based or bottom-up planning, the agent “discovers” the resource; it is already acquired. Thus, opportunistic planning should have a higher chance of success simply because the resource acquisition step has already been accomplished.
This higher likelihood of success for resource-based plans is not limited to opportunistic plans. Wilensky (1978) discusses the idea of goal subsumption for situations in which a planner takes one action to satisfy a recurring goal. For example, a man who tires of going to single bars in service of his S-SEX goal may choose to get married. A person may get a job to provide a steady income in service of the goal to acquire money. Someone who drives an hour to work each day may decide to move closer to the job, thus cutting down on the commuting time.
These actions — getting married, getting a job, and moving — are examples of what Wilensky termed goal subsumption. In each case, the agent has a persistent goal with a corresponding plan for achieving that goal, but adopts a new plan to avoid a negative side-effect from the old plan. In our current framework, we observe that the new plan involves the acquisition of a persistent resource, namely, a wife, a job, and a house in a better location. The resource enables plans that satisfy the persistent goal in a more reliable and effective manner.
People commonly acquire such persistent resources. A wallet or a pocket book provides a convenient way of satisfying the enabling conditions a large number of plans — both expected and unexpected. A wallet or purse might typically contain a variety of resources, such as the following.
Money — both cash and credit cards.
Credentials — driver permit, check cashing card, work ID.
Facts — addresses, phone numbers, passwords, appointments.
Functional Objects — keys.
By carrying a wallet or purse, the agent is better able to pursue plans that require those resources. Note: today a smart phone serves many of these purposes.
Our examples of resources have been fairly tangible commodities, such as money. This analysis is similar to other planning models [Schank and Abelson, 1977, Schank, 1986, Wilensky, 1978, Wilensky, 1983]. We can also consider cognitive processes as resources in a similar way, and observe the implications for the importance of goals.
One may view the cognitive process that manipulates goals as a type of scheduler. There are numerous tasks or goals to pursue and the scheduler decides which ones to activate. The process is roughly as follows.
Identify active task with highest priority.
Execute that task.
During execution, service interrupts of higher priority than current task.
This process has several aspects that make it more complex:
• A task may have several different priorities relative to different higher level goals.
• A task may cause other tasks to be reordered.
• A task may simply be to determine the priority of another task.
Measuring the priority of active tasks is not necessarily simple to do. One method would be to rank the tasks by importance, but this could be problematic due to the resource disparities of tasks. Generally, as important tasks become blocked due to resource limitations, lesser tasks are pursued. Another approach might be to rank active tasks inversely by resource requirements, allowing the easy tasks to get done first. Moreover, the importance of a task might be affected by how many resources have already been allocated for that task. A person who has “invested” in a task may be more inclined to pursue that task so as not to lose the investment.
Interrupts are meant to handle inputs not directly related to the current process. For example, an interrupt occurs when one is reading an article and hears the phone ring. The current process is reading. An extraneous input — the telephone bell — causes the interrupt mechanism to fire. The scheduler decides whether the priority of answering the phone is high enough to preempt reading. This decision is potentially complex. The following examples illustrate these notions.
Example 4.30 John was reading the newspaper when the phone rang. It was a wrong number. John finished the sports page.
Example 4.31 John was reading the newspaper when the phone rang. His wife had had an accident and was at the hospital. John ran out of the house.
Example 4.32 John was defusing the bomb when the phone rang. John figured they would call back later.
In the first two examples above, John is pursuing a fairly low-level goal when the phone rings. This interrupt is deemed of higher importance, so John services it — he answers the phone. In the second case, the phone call causes John to develop a goal with a higher priority than reading the newspaper, namely, going to the hospital to see his wife.
In the third case, answering the phone is much less important than defusing the bomb. That is, if he serviced the interrupt, he might fail to achieve the higher goal of preventing the bomb from exploding.
This scheduler model embodies the notion of busy. To say that a person is busy means that the person can service only the highest level interrupts. Previously, we considered this property an aspect of the time resource. This idea can be applied to inferences about goal importance.
Example 4.33 “I’m never too busy to talk with a constituent,” said Congressman Brown.
Example 4.34 After his promotion, John was often too busy to play golf on Wednesdays.
Another notion brought out by this model is the work/play dichotomy of goals. Work goals are serious goals, and play goals are frivolous goals. As the last example shows, goals that are being pursued seriously (work goals) are seldom interrupted by frivolous goals (play goals).
Seriousness relates to the manner in which a task is executed, and the degree to which goal satisfaction or success is required. Work tasks are very much oriented to goal achievement (A-goals), whereas play tasks involve less tangible satisfaction (E-goals). The failure consequences of serious goals have a negative effect on important goals. Frivolous goals have negligible failure consequences.
The three main cognitive processes we wish to examine are processing time or attention, memory, and learning. Attention translates fairly directly as a quantifiable resource. The amount of processing time that the achievement of a goal may require generally correlates with the goal’s importance.
Example 4.35 John spent weeks thinking about his campaign strategy.
Example 4.36 John did not pay attention to his campaign’s letterhead design.
In the first example, we can infer that John’s campaign strategy was very important to him. In the second case, we can assume that John did not care about the campaign stationery. It is reasonable to treat processing time or attention as a measure of importance. Thus, attention, as a resource, provides a method for inferring the relative ranking of goals.
Principle 3 The importance of a goal is proportional to the processing time which the agent is willing to expend in pursuit of that goal.
Processing time is not strictly the same as the time resource discussed earlier.
Example 4.37 While driving to work, John was thinking of what to get Mary for her birthday.
Example 4.38 While driving to work, John was listening to the talk show on the radio.
Example 4.39 While driving to work, John was day-dreaming about his upcoming vacation.
Example 4.40 While driving to work, John was planning the bank robbery.
In each of these examples, John will presumably get to work no sooner or later than if he had been devoting all of his attention to the task of driving. Some tasks take the same amount of time no matter how much thought a person applies to the task. By comparison, consider the case of a time-shared computer in which CPU time is analogous to attention, and “wall clock” time is equivalent to our typical time resource. Some computer programs will of course finish sooner if more CPU time is devoted to them, however, more CPU time will not necessarily speed up the printing of a document or input from a terminal. The former processes are CPU-bound, and the latter are I/O-bound.
The question of allocation of attention has been addressed by Schank (1978) who proposed interestingness as a method for controlling inferences during language understanding. “When something is deemed to be of interest, it is tracked by the reader. That is, we pay attention to it. What does attention mean? It means that we let our inference processes loose.” [Schank, 1978]
We view the present discussion of importance as an elaboration of the interestingness concept. Specifically, we see that the ranking of goals according to importance provides a natural organizing principle for the various categories of interesting things, such as the list proposed by Schank: DISEASE, POWER, SEX, MONEY, DESTRUCTION, ROMANCE, and DEATH.
Lebowitz’ program IPP [Lebowitz, 1980] used interestingness to guide processing in a manner analogous to the goal-based approach to attention suggested here.
One specific task requiring attention is planning. We suggest that the processing time devoted to planning will reflect the underlying importance of the goals involved. For example, one part of the planning process is to determine how the proposed plan will affect the agent’s goals. Some goals will be positively affected and others may be negatively affected, either directly or through side-effects. The planner needs to know what goals will be achieved and what goals will be blocked. This process of evaluating the outcome of a plan should reflect the underlying importance of the agent’s goals. As we have argued, an agent has a vast multitude of goals, and a given plan might affect any number of them. It would be unreasonable to propose a planner that checks the plan side-effects against all goals. However, it is computationally feasible to pay attention in a top-down fashion to high-level goals when evaluating a plan outcome. This mechanism would give the planner an opportunity to anticipate plan side-effects relating to important goals.
We aver that the structure of memory should mirror mental processing, and vice versa. Our memory should be a reflection and an artifact of our cognitive processing. The way in which we remember an event should be determined in part by the way in which we originally thought about the event. The act of remembering should alter the structure of memory. Attention and memory are complementary phenomena. Table 4.2 characterizes attention and memory using our postulated resource dimensions.
In contrast to attention, the notion of memory as a resource is not so straightforward. We do not view memory to be an incremental resource which is expended. Memory is neither created nor destroyed. We do not propose that one can either increase or deplete one’s memory. What does vary in memory is the speed or ease with which information can be retrieved. The proper measurement of the memory resource is then how accessible is the relevant information for the achievement of the goal.
Principle 4 The importance of a goal is proportional to the ease with which information relating to the achievement of that goal can be remembered.
Implicit in this claim is the notion that there are various levels of indexing and organization in memory, and that these levels are reflected in the goal hierarchy. One very straightforward model to account for this would be that memory search is keyed directly off the goal hierarchy. Memories of episodes involving plans, and resources would be indexed through goals, and the goal hierarchy with its related memories are then searched in a top-down fashion. This would allow the relevant memory relating to the most important goal to be found first. Moreover, the processing time allocated for memory search would be limited by the importance of the related goal. Thus, information that is very important would be likely both to be stored in an accessible location, and to be sought out more thoroughly. Less important information would be stored more remotely, and would have a less diligent search effort. These factors converge to suggest that important information will be found, while trivia will not.
Note that importance might be construed differently at storage and retrieval times. This difference becomes a problem if an event was considered unimportant, but later gains significance. Such an event would be harder to recall than one to which greater attention was paid initially. The opposite situation, in which a once important event becomes less critical, does not pose a retrieval problem, since there is little interest in remembering the event at all.
Example 4.41 At the party, John forgot the name of his boss’s wife.
Example 4.42 John could not remember the name of his third grade teacher.
In the first example above, we assume that John had not made an effort to learn the name of his boss’s wife, assuming that he would not have any contact with her. However, it later became important for him to remember the name. In the second case, we assume that when John was in third grade, he knew the name of his teacher. However, that information is no longer important for him to know.
Several heuristics fall out of this model. First, people are likely to remember things which are important to them and likely to forget things which are not. This accounts for why a wife might get angry when her husband forgets their anniversary, or why people are likely to remember the first time they had sex.
Second, people are able to use importance criteria to judge what they are likely to know and, similarly, what they are likely not to know. Consider two examples.
Example 4.43 John did not know the population of Mali.
Example 4.44 Reagan did not know if he had approved the sale of arms to Iran.
We are inclined to believe the first example. We can assume that it is not important for John to know the population of Mali. However, we find it difficult to believe the second example, since approving arms sales to Iran would have violated a fundamental administration policy. Presumably that policy would have been considered important.
In the preceding chapter, we discussed knowledge and mental skills as types of resources. Memory is the repository of knowledge and mental skills. Our claims about the role of importance in organizing memory apply to knowledge and skills which are the content of memory. We recognize that memory is not like a homogeneous computer data base. The contents of memory are quite varied, as the following categories suggest.
• Durable information. Knowledge that remains accurate over time: the value of pi, your name, the capital of France, the number to call for directory assistance.
• Transient information. Knowledge about things that change in the world: what you had for lunch, where you parked your car, who is picking up the children at school today, the name of the former Soviet Union.
• Plans. Knowledge about procedures for achieving goals: how to call someone on the phone, how to get from your office to your house, how to print out a file.
• Cases. Episodic knowledge about past experiences: the first time you ate at a three-star restaurant, what happened at lunch today, the time you totalled your Dad’s new car.
We claim that information, plans, and cases are all cognitive resources that reflect the importance of the underlying goals. Their availability in memory will be proportional to the priority of the goals to which they relate.
In modeling memory phenomena, we are primarily concerned with episodic memory [Tulving, 1972]. Goals have played a role in previous reconstructive models of episodic memory [Kolodner, 1980, Schank, 1982]. One particularly relevant study is that of Anderson and Pichert (1977), which showed a change in a processing goal producing a significant change in memory retrieval. Our model is consistent with those results.
There are clearly episodic elements in other types of memory. For example, knowledge encoded in a semantic memory framework might have an index to the episode during which that information was learned [Schank, 1975b]. This brings us to our next subject, learning.
Importance plays a leading role in learning. The simple view is merely that one learns better if one is motivated to learn. People can more easily learn things that are relevant to their lives — and thereby relevant to their personal goals. The more motivation, the better. The more relevance, the better.
We can examine the relation between learning and goal importance a bit more analytically though. We will first posit a principle concerning learning.
Principle 5 Learning is a cognitive process that requires attention and memory resources.
When we learn something, we are engaged in a process affected by both processing time and memory organization. Combining the preceding three principles, we may conclude the following.
Principle 6 The importance of a goal is proportional to the ease with which a concept relating to that goal can be learned.
The eager student will learn. The bored student will not. However, motivation is not sufficient. Learning will vary among students according to ability as well. Still, learning is explicitly tied to memory processing, which we have argued is directly affected by the importance of goals.
Schank has pointed out the role that explanation plays in learning [Schank, 1986], and noted that what items tend to be the subject of explanation are determined by a “spectrum of need.” We suggest that the goal hierarchy exactly reflects this subjective spectrum. The phenomena that we choose to try to explain are those that are most important. Clearly, there are innumerable phenomena that occur daily, but we focus our attention on but a small portion of those. We suggest that importance is a crucial part of deciding what to explain.
Furthermore, even after we have decided what we wish to explain, importance again plays a role in the nature of the explanations themselves. Schank has noted different types of explanation are needed for different phenomena. Broadly speaking, there are physical phenomena and intentional phenomena. The former would include events such as the weather, and disease etiology. The latter would include much of human behavior, as well as functional artifacts and institutions, such as television sets and governments. Our model of planning has little bearing on explanations of physical phenomena. However, we can make several observations concerning explanation of intentional behavior.
First, we note that the process of explanation is triggered by an expectation failure. Something occurred which we did not expect. We detected an anomaly. In the present planning model, such an anomaly could occur in a variety of circumstances, such as the following:
• The agent is pursuing a plan that is not instrumental to any known goal of the agent.
Example 4.45 John painted his face green.
Explanation: Infer that the agent has some goal for which this action is instrumental, e.g., going to a costume party.
• The agent pursues a goal that is in violation of a known goal, or the agent fails to pursue an action to satisfy a known goal.
Example 4.46 John gave away all his worldly possessions.
Example 4.47 John refused to take his medicine.
Explanation: Infer that the agent has abandoned the known goal, or is pursuing a higher-level goal that is in conflict, e.g., John has taken vows of poverty, and John is allergic to the medicine.
How do we decide if an explanation is plausible?
Example 4.48 John cut off his right arm because it had gangrene.
Example 4.49 John cut off his right arm to lose weight.
The former explanation is plausible. The amputation of the arm might prevent more serious consequences. The latter explanation is curious. Presumably, the consequences of being overweight are not severe enough to merit such drastic steps, since there are alternative ways of losing weight. Importance is a key factor in judging plausibility.
One common type of intentional explanation is the excuse. Excuses are a kind of explanation, the plausibility of which can depend on importance.
Example 4.50 Mary could not go out with John because she had the flu.
Example 4.51 Mary could not go out with John because she had to study.
Example 4.52 Mary could not go out with John because she had a hangnail.
Example 4.53 Mary could not go out with John because she had to wash her hair.
It is not hard to recognize that that the first two excuses above are more compelling than the last two. How are we able to arrive at this judgment? In the first case, we understand that having the flu will impair Mary’s health so that she cannot pursue certain activities, and that furthermore, her health might deteriorate if she does not take care of herself. Presumably, going out with John is not more pressing or important than her immediate health needs. At the same time, we could easily interpret the following case.
Example 4.54 Mary went to the Michael Jackson concert even though she had the flu.
Here the Michael Jackson concert is considered of greater importance than Mary’s immediate health.
It is also possible for an explanation to be implausible for other reasons.
Example 4.55 Mary could not go out with John because she was abducted by Martians.
Rejecting this explanation is based on normative knowledge, rather than Mary’s priorities. Thus, we do not claim that importance is a sufficient criterion for judging plausibility. We are suggesting that importance will often play a role in such judgments. Our analysis here is intended to be consistent with the related work on explanation by Kass (1990) and Leake (1990).
In our taxonomy and analysis of goals and resources, we identified a dimension of persistence. Some goals and resources last longer than others. Goal persistence is reflected in persistent affective states, or moods. We can call a short-term or acute affective state an emotion, in contrast to long-term or chronic moods.
Consistent with our model, we suggest that moods may result from chronic high-priority goals. The intensity of the mood will reflect the importance of the related goals. We look at two basic moods: good and bad.
We examine two relevant components of moods — their origin and their behavioral results. First, how does someone get in a good mood? Someone is in a good mood if he achieves (or expects to achieve) a persistent, regnant goal.
Example 4.62 After a long, hard mayoralty campaign, John won the election.
Example 4.63 At the shopping mall, John found a parking place near the entrance.
In the first example, John has achieved a persistent goal — he has been elected mayor. We infer that John would be in a good mood at this point. In the second example, John has achieved a goal, but one of lesser importance and of low persistence. We doubt that John’s mood will be much affected by his good parking place.
In determining a mood, we must identify the regnant persistent goal. It is not enough that an agent achieve an important persistent goal. That goal must not be eclipsed by another goal.
Example 4.64 After John won the election, he discovered that he had cancer.
Presumably John’s good mood would fade away at this news.
The second question is: What is a person likely to do when in a good mood? What are the behavioral consequences of a good mood? We can use the principle of importance to address this question. An agent is in a good mood if he has achieved an important goal. According to the principle of importance, an important goal is one to which an agent is willing to allocate considerable resources to achieve. Thus, before achieving the important goal, the agent was prepared to expend additional resources toward the pursuit of that goal. In a sense, the agent had reserved or committed resources for that goal. Once the goal is achieved, those reserved resources are no longer committed. The agent is able to allocate those previously committed resources to other tasks. Thus, an agent who is in a good mood effectively has excess resources that can be used for new goals.
There is a heuristic that suggests that you get someone in a good mood before delivering bad news or making a request. For example, the wife cooks the husband a big dinner before telling him about the auto accident. The employee waits until the boss is in a good mood before asking for a raise. Somehow, bad news does not seem so bad when a person is in a good mood. We suggest that our goal-based model partially accounts for that phenomenon. A person in a good mood perceives that extra resources are available for handling new goals. A problem, in the form of bad news, is a type of new goal. Thus, the bad news will seem more tractable. Similarly, a request is also a new goal. The extra resources can go toward satisfying the new request.
The model for a bad mood is analogous to the good mood model. A bad mood arises from the failure to achieve an important persistent goal.
Example 4.65 John lost the election.
Example 4.66 John did not get the Henderson contract.
Example 4.67 John’s wife left him.
Example 4.68 John found out the tumor was malignant.
In each case, the agent fails to achieve or maintain a long-term important goal. One consequence is a loss of resources. The agent must reallocate existing resources to make up the difference which means that some of his earlier goals will now go unsatisfied. An agent in a bad mood will thus have to sacrifice goals that he had previously pursued. This means that the agent will be much less open to the pursuit of any new goals.
In the next chapter we discuss interpersonal relations. We shall extend the analysis of emotions and moods to cover adopted goals. Thus, when something very good happens to your friend, you may also feel happy — depending on the importance of the event and the relationship. By adopting the goals of your friend, you identify with him and mirror his emotions. This is an empathic view of emotions, and can be used for emotional reactions to fictional characters, such as in a book or movie. If you identify strongly with the character, you can vicariously experience the emotions of the character.
In closing we should point out that our model does not posit a force to emotions per se. We do not suggest that anger or rage, for example, are themselves theoretical entities that affect an agent’s behavior. Rather, we argue that emotions are epiphenomenal. Emotions are a reflection of the underlying goal hierarchy, which is the primary motivational force. Emotions serve to communicate an agent’s goal state to himself and others. However, we do not require that emotions are themselves motivational.
• We define a resource as that which satisfies an enabling condition of a plan.
• The principle of importance states that resource allocation reflects the importance of a goal.
• The importance of resources is a subjective function of an individual’s goals and available resources.
• The principle of investment states that the importance of an active goal will reflect the resources that have already been allocated to that goal.
• We present a taxonomy of fifteen types of resources.
• We analyze this classification along eight dimensions.
• The comparison of resources is subjective and will reflect the importance of an agent’s underlying goals.
• Just as we infer default goals for different types of people, so may we infer default resources based on limited knowledge of a person.
• Goal development may be top-down (goal-based), or bottom-up (resource-based). The latter type of goal development is opportunistic, and generally has a greater chance of success.
• Goal subsumption involves the acquisition of a persistent resource to satisfy a persistent goal.
• The cognitive phenomena of attention, memory, and learning can be treated as resources to be allocated by a planner.
• The importance of a goal is proportional to the processing time which the agent is willing to expend in pursuit of that goal.
• The importance of a goal is proportional to the ease with which information relating to the achievement of that goal can be remembered.
• Learning is a cognitive process that requires attention and memory.
• The importance of a goal is proportional to the ease with which a concept relating to that goal can be learned.
• The process of explanation as a part of learning reflects the importance of related goals both for detecting anomalies to be explained and determining the plausibility of explanations.
• The importance of a goal is proportional to the degree of affective response to the status of that goal.
• When an event affects several goals at once, the emotional response will reflect the status of the regnant goal.
• Goal persistence is reflected in chronic affective states, or moods. The intensity of the mood mirrors the importance of the affected persistent goal.
• Our model suggests that emotions are an epiphenomenal reflection of the goal hierarchy, rather than a motivating force themselves.