<< Click to Display Table of Contents >> Navigation:  Decision-theoretic modeling > Utility

An integral element of all decision problems, one without which no decision can be made, is the notion of preference. Very often, preference can be based on an objective quantity, such as material usage, factory output, or financial gain. Typically, however, decision problems involve quantities that have no obvious numerical measure, such as state of health, customer satisfaction, or pain. Another complication is a possibly conflicting set of attributes, such as price and quality. Even if a numerical measure of goodness of an outcome is available, such as is the case with financial gains and losses, it may not reflect well decision maker's preferences in presence of risk.

Decision theory introduces a measure of preference, known as utility. Utility is a function mapping the attributes of the possible outcomes of a decision process on the set of real numbers. Utility is determined up to a linear transformation, i.e., a decision maker's preference over different decision alternatives is invariant to multiplying the utility by a non-negative number and adding a constant. This implies that utility has neither a meaningful zero point, nor a meaningful scale.

Utility is by assumption subjective: various decision makers facing the same choice and even sharing the same set of beliefs about the world may choose differently because of their different preference structure and different utility functions. A utility function for any decision problem needs to be obtained from a decision maker. The process of obtaining a utility function from a decision maker is known as utility elicitation.

It is worth pointing out that variables measuring utility are always continuous: they can assume any values from a continuous interval. Sometimes they are mistakenly taken for discrete variables, as in graphical models, such as influence diagrams, they usually have discrete parents and take a finite number of values. It is much clearer to see that multi-attribute utility (MAU) variables are continuous - they specify a function by which the values of their parents, utility nodes, are combined.