Knowledge Content Library
How to Value Imperfect Information
Presenter: Ron Allred and Jon Anker, ConocoPhillips
Presented at the 2003 DAAG Conference in Houston, Texas.
Difficulties associated with decision-making under uncertainty are
familiar to most people. Companies and individuals regularly make decisions where important information
about factors that could significantly affect the outcome of decisions is
lacking. Two general patterns with
regards to decision-making can be recognized.
- A general EMV pattern, where the decisions occur up
front and then all the uncertainties occur after those decisions are made.
- A phased decision pattern, where the decisions are
interspersed with the uncertainties.
A phased decision pattern is indicative
of a value of information situation.
A value of information situations requires the assessment of whether
the acquisition of information would affect future decisions. For most circumstances the information to be
collected is not perfect, the prediction may be wrong. Hopefully once information has
been acquired a better perception about the likelihood’s of the states of
nature will exist, but because the information is imperfect we will not have
complete information about the actual state of nature.
When evaluating the viability of acquiring information, the most
straightforward approach is to use decision tree analysis to determine whether
the EMV of the project is greater than or equal to the project without acquired
information. The normal steps to take in
this type of analysis would be to;
1. Establish actual(prior) probabilities. These are the
actual probabilities for some event prior to collecting information.
2. Establish indicated(conditional) probabilities. These are
the probabilities indicated by the acquired information if an actual event
3. Determine the posteriorprobabilities using Bayes’ Theorem.
Determining the probabilities of the outcome of an actual (with prior
probability) event following the acquisition of information (indicated
4. Comparison of
EMV’s with and without the acquired information.
Conceptually we can solve most EMV problems using decision trees,
however if more and more uncertainty events are added to a decision tree model
then the sudden increase in the number of branches makes calculations both
cumbersome and impractical. It has now
become routine for EMV calculations to be done using electronic spreadsheet
programs based on "Monte Carlo” randomised simulation. These programs use a random sampling process
to generate inputs from uncertainty distributions, which are then processed by
a mathematical model.
While Monte Carlo simulation programs are easy to use
and can deal with a large number of uncertainty events, they are not well
suited for use in value of information situations. There is no difficulty in constructing
probability distributions for actual (prior) and indicated (conditional)
events, but it is very difficult to construct a resultant posteriorprobability distribution.
Based on simulation work concerning the value of
appraisal drilling information prior to field development, we have devised a
methodology that allows for determining the value of information using Monte
Carlo sampling techniques. This
methodology is based on comparisons of indicated (conditional) probabilities
with compressed ranges associated with actual (prior) probability
Keywords: Value of information voivoc valoi, modeling modtree, decision trees dectree, simulation mcsim