Chapter 1
Goals and Decisions
Chapter Summary
1.1 Introduction
A realistic model of decision making must account for a multitude of goals and limited resources. Many problems will not admit perfect solutions. Practical decisions often involve trade-offs. Examples of such choices include buying a car or a house, prescribing a medical treatment, purchasing computer hardware and software, or getting married.
Standard prescriptive models of decision analysis [Raiffa, 1968] maximize an expected value based on the probabilities and payoffs of different possible outcomes. We propose an alternative descriptive model of decision making based on an analysis of an agentâs goals, past behavior, and relationships with other agents. Our model suggests that a decision maker has a multitude of goals beyond simply maximizing payoffs and minimizing risks. The agentâs goals include both personal goals and adopted goals derived from interpersonal relationships. The agent must resolve goal conflicts by making trade-offs.
In this goal-based model of decision making, the relative priorities of goals drive the decision process. This model serves as the basis for the VOTE program which simulates Congressional roll call voting decisions, based on the relative priorities of goals and constituent relationships, as well as the need to provide an explanation for the decision.
In developing the role of goals and priorities in decision making, we view goals not as atomic entities, but as composite objects with a defined structure. Goals are complex. We explore a number of significant dimensions of goals, and develop a broad range of types of goals that share basic properties relevant for decision making.
1.2 Ubiquitous Goals
Goals are a pervasive element of our language and thought. Consider the following excerpts from articles in the New York Times of September 21, 1990.
The above quotations provide explicit and implicit statements of goals. The following stories involve other aspects of goal pursuit, such as projections of the likelihood of achieving a goal, or the need to modify a goal or a plan.
The following three stories illustrate the way in which a decision may require justification to the degree that its consequences violate goals.
As these examples suggest, through both explicit and implicit reference, we commonly speak of goals, plans, motives, responsibilities, desires, intentions, wants, needs, and many other goal-related concepts. The pursuit of goals, and the concomitant communication, discussion, and evaluation of goals, are prominent parts of human behavior.
Goals may be a ubiquitous phenomenon, yet goals remain abstract. We cannot see a goal. Goals are intangible.
Artificial intelligence (AI) has long recognized the utility of goals in the simulation of human planning and problem solving. Traditional AI models have generally treated goals as uniform, homogeneous constructs. Such goals may be achieved through the execution of plans which are themselves composed of sub-goals. GPS [Newell and Simon, 1963] and the blocks world programs [Winograd, 1972, Sussman, 1975] implicitly accepted the view that a goal was equivalent to a specific desired state of the world. A goal was something to be achieved through some set of actions.
Doyle (1990) recently challenged this view of homogeneous goals. Doyle cites Newellâs Principle of Rationality:
Doyle stated that âthis principle ignores comparisons among goals and among the different methods for [achieving] goals,â and goes on to endorse an economic theory of choice.
Another perspective on goals was offered by Schank and Abelson (1977). For story understanding tasks, they proposed a broader view of goals which distinguished among a variety of goal types including goals of satisfaction, enjoyment, achievement, preservation, crisis, and instrument. The Schank and Abelson taxonomy was suggestive of the underlying properties of goals.
We propose an extension of the Schank and Abelson view of goals, embracing the following general points.
Our task is to begin to characterize and analyze goals. We shall try to ascertain or propose an underlying structure for goals. We shall use this structure to develop a notation for goals, and to provide an explicit representation for goals. This representation may then serve as the basis for a computational model of intentional behavior, specifically, decision making.
We should also mention that we do not intend to produce a formal semantics for goals. Some AI researchers believe that the way to make a program smart is to provide it with a formal semantics for its knowledge. This notion seems analogous to a belief that the way to become rich is to open a checking account. Being able to write checks and get monthly statements may provide convenient transactions and bookkeeping, but a checking account is neither a necessary nor sufficient condition for wealth. In fact, there are many types of wealth that would not be reflected directly in a checking account, such as real estate, jewelry, or a stock portfolio.
Similarly, we believe that formalizing knowledge representations is neither a necessary nor sufficient condition for endowing computer programs with intelligence. Defining wealth in terms of the balance in a checking account may well ignore many less liquid assets. Formalizing knowledge representations may well ignore less formal types of knowledge.
The time to get a checking account is after you have earned some money. The time to formalize knowledge is after you have acquired some knowledge.
1.3 Chemistry, Music, and Goals
A goal is an abstraction. You cannot see or touch a goal, just as you cannot see or touch gravity. Physicists have postulated forces like gravity to explain the physical world. Gravity is not directly observable, but its effects are. Gravity is a useful theoretical construct for explaining physical phenomena. Goals can play an analogous role in the explanation of human behavior. Gravity is useful in predicting the fall of a leaf from a tree or the motion of planets through space. We suggest that a theory of goals is useful in modeling and explaining individual and interpersonal behavior.
In outlining a theory of goals and decision making, we shall engage in observation and analysis. Across a wide range of disciplines, we note the process of observation, categorization, reduction, and notation. In suggesting that we focus on goals as a topic for classification and analysis, we first offer analogies to other well-established reductive enterprises, namely, the chemical theory of elements, and the development of musical notation.
The kinetic theory of atoms can be traced back to Demokritos (c. 460â370 B.C.) who proposed four basic elements: earth, air, water, and fire. Earlier philosophers, such as Heraklitos, had proposed related theories focusing on but a single âfirst sourceâ of matter, such as fire.
Over the centuries, the classification of elements expanded and became more refined. Chemists observed standard ways in which elements would combine to form more complex substances. They discovered that the same basic substance, such as water, could occur as a gas, liquid, or solid. Chemists noted standard properties exhibited by the elements themselves, such as atomic weight, valence, specific gravity, boiling point, and density. They developed a systematic notation for chemical elements and compounds.
In 1869, the Russian chemist Dimitri MendelĂ©ev proposed a classification system for the elements that emphasized the cyclical reoccurence of the elementsâ chemical properties with increasing atomic weight. Under MendelĂ©evâs system, elements fell naturally into eight distinct groups. Elements in each group exhibited similar chemical properties. MendelĂ©ev was able accurately to predict the properties of three elements that had not yet been discovered, but which would naturally occur at gaps in his table.
Chemists and physicists were later to discover explanations for the periodic properties in the electron shells of the elements. MendelĂ©evâs periodic table of elements reflected the fundamental physical atomic structure of matter.
The periodic table and the related chemical notation are abstract concepts that describe and depict natural physical phenomena. Over the centuries, the scientific enterprise of chemistry has refined and expanded our understanding of the underlying properties and physical structure of matter. By analogy, we suggest that a study of intentionality may follow a similar path. We first observe the role of goals in human behavior, and develop taxonomies for classifying different types of goals. We try to discern properties of goals that may lead us to discover an underlying regular structure.
The field of music provides another domain in which observations of a physical phenomenon led to a system of classification and notation. As with chemistry, the Greeks appear to have started the investigation. Early mathematicians, such as Pythagoras, observed the regular relationships between the length of a vibrating string and its pitch.
As western music developed, so did the notation for expressing the underlying atomic unit of music: the note. Just as a chemical element is defined by specific structural properties, a note di...