Stereotype error

One less than obvious but important factor in stereotype models is error. For example, a stereotype belief variable might take its value from ten dialogues, in five of which the value was true and in five of which the value was false. While the stereotype value would then be 0.5, the true value would be normally distributed around 0.5. This has the consequence that decisions that depend on the value of this variable may in some cases be in error. For example, there may be what will be termed a decision surface between two alternatives that occurs at the value of 0.4 for the variable. The decision surface is a surface in belief space across which the maximum utility alternative changes. Although the variable is estimated at 0.5, the actual value falls below 0.4 with a certain probability, resulting in a mixture rather than just one of the alternatives being taken. To deal with this problem, a sampling system has been implemented which randomly varies each value to simulate the error. Using 1000 samples, the system returns the alternative of maximum utility over the 1000 samples. 1000 samples is enough to ensure statistical significance of the decision, unless the alternatives are very close in utility. An example of use of the sampler will be given in the next chapter.