Intelligent decision support was the topic of week 8’s lecture. This is a topic that has been built on the ‘Fundamental Preference Assumption’ (given choices A, B either A > B, B> A or A ~ B). This topic is closely related to our previous lectures which centered around reasoning with uncertainty.

Rational preferences are a prerequisite for intelligent decisions. Characteristics of rational preferences are:

  • Orderability
  • Transitivity
  • Continuity
  • Substitutability
  • Monotonicity

Mapping of preferences that may not have a readily comparable outcome is achieved through utility values. We did not cover any material on the development of utility values. I believe that to increase the success rate of intelligent decision systems, collaboration between end users and implementors must be conducted. A perfect systems with poorly represented utility values will fail.

Principle of Maximum Expected Utility (MEU) – An agent is rational iff it makes decisions that reflect MEU (I would argue rather that ‘An agent can’t be rational if it does not make decision based on MEU). Rationality should encompass consideration of the source of utility values.

Using Bayesian networks with decision and utility nodes, Dynamic Utility Networks can be developed. Depending on the information available, the maximum expected utility of decision can be calculated. This is a key concept for rational planning in uncertain environments. The value of new information can also be calculated using Shannon’s utility gain, a topic to be discussed next lecture.

ExpectedUtility
Expected utility in uncertain environments is linked with Bayes theorm