1.5 Agent Design Space

1.5.7 Preference

Agents normally act to have better outcomes. The only reason to choose one action over another is because the preferred action leads to more desirable outcomes.

An agent may have a simple goal, which is a proposition the agent wants to be true in a final state. For example, the goal of getting Sam coffee means the agent wants to reach a state where Sam has coffee. Other agents may have more complex preferences. For example, a medical doctor may be expected to take into account suffering, life expectancy, quality of life, monetary costs (for the patient, the doctor, and society), the ability to justify decisions in case of a lawsuit, and many other desiderata. The doctor must trade these considerations off when they conflict, as they invariably do.

The preference dimension considers whether the agent has goals or richer preferences:

  • A goal is either an achievement goal, which is a proposition to be true in some final state, or a maintenance goal, a proposition that must be true in all visited states. For example, the goals for a robot may be to deliver a cup of coffee and a banana to Sam, and not to make a mess or hurt anyone.

  • Complex preferences involve trade-offs among the desirability of various outcomes, perhaps at different times. An ordinal preference is where only the ordering of the preferences is important. A cardinal preference is where the magnitude of the values matters. For example, an ordinal preference may be that Sam prefers cappuccino over black coffee and prefers black coffee over tea. A cardinal preference may give a trade-off between the wait time and the type of beverage, and a mess versus taste trade-off, where Sam is prepared to put up with more mess in the preparation of the coffee if the taste of the coffee is exceptionally good.

Goals are considered in Chapter 6. Complex preferences are considered in Chapter 9.