BoundedRationality_TheAdaptiveToolbox.pdf

112 Gary Klein
selecting the best one will decrease. (A similar issue was raised by von
Winterfeldt and Edwards [1986] in their discussion of the flat maximum.) Thus,
the comparisons that mean the least (for options that are microscopically different
in value) will call forth the greatest effort. This argues that optimization may
exist primarily for local tasks and may be unworkable when more global consid
erations such as effort are taken into consideration. To the extent that this is true,
it then diminishes the relevance of optimization for field settings and raises other
difficulties, such as the self-referential problem of calculating the cost of calculating.
Thus, if I want to optimize, I must also determine the effort it will take me
to optimize; however, the subtask of determining this effort will itself take effort,
and so forth into the tangle that self-referential activities create.
7. The options must be thoroughly compared to each other. This entails
comparing each option to every other option and making the comparisons on a
full set of dimensions. Janis and Mann (1977) contrasted optimizing with
satisficing and noted that satisficing permits minimal comparisons. The decision
maker does not have to compare every option to every other option. Optimizing
requires that all the reasonable options be compared on all dimensions,
in carrying out a multiattribute utility analysis.
Janis and Mann (1977) further stated that a satisficing strategy is attractive
because it can be conducted using a limited set of requirements. They point out
that an optimizing strategy contains an unrealistically large number of evaluation
criteria and factors that have to be taken into account. Thus, the comparisons
must be conducted over a wide range of considerations. This was discussed
earlier in terms of the need to look at global issues. Hill-climbing algorithms represent
a middle ground between satisficing (e.g., selecting the first option that
works) and exhaustive comparisons (all options on each dimension), but they require
a well-structured problem space and do not constitute optimizing.
8. The decision maker must use a compensatory strategy. In critiquing the
feasibility of an optimization strategy, Janis and Mann (1977) note that the comparison
needs to take into account the degree of advantage and disadvantage that
each option carries, on each evaluation dimension, rather than merely noting
whether or not a difference exists. A problem arises when we consider findings
such as Erev et al. (1993), who found that the decomposition needed to conduct
formal analyses seemed to result in lower quality decisions than an overall preference
would yield. If the implementation of optimization procedures interferes
with expertise, then the methodology may have a serious disadvantage.
9. The probability estimates must be coherent and accurate. It makes little
sense to use the appropriate methods if those methods are not carried out well.
Beyth-Marom et al. (1991) noted that in carrying out a decision analysis, the
probability estimates for different outcomes must be reasonably accurate. This
requirement makes sense. Yet how should a decision maker determine if the requirement
is satisfied? Unless the different outcomes are systematically tested,
the requirement is more an admonition to try hard to be accurate. Sometimes,