Stereotypes are most useful when a system is used only once by a particular user, yet the belief states of the user population form tight clusters from which classes can be derived, and when it is easy for the system to acquire many samples of data from the population. Stereotype acquisition is a matter of finding the belief model which maximises the probability of the belief states of the members of the stereotype class, given that belief model. This is a case of "maximum likelihood estimation", where a model is required that predicts the data in the distribution that was observed.
Only one stereotype is used here, representing all users, avoiding the problem of retrieving the correct stereotype for a given user. Stereotypes can also be used to represent agents whose beliefs vary from time to time, or are influenced by outside factors that cannot be modelled. The stereotype then is representative of the expected beliefs of the single agent, given a number of samples of that agent's dialogues.
In some cases where the system is using a stereotype model of a user, each user drawn from the stereotype interacts with the system only once. An example would be a once-off travel booking. In this instance the stereotype model is only one level deep, since the user has no experience of the system, and therefore does not form a dynamic stereotype over time. On the other hand, if the same system interacts with the same user many times, the system must adapt the model at the second and subsequent levels, as each develops a dynamic stereotype of the other.
Acquisition of stereotype models is straightforward. A set of example dialogues is taken, and the belief revision mechanism is used to update the beliefs. For each belief in the belief state, a mean value over the example dialogues is obtained to compute the stereotype value. From this mean value, the correct distribution of belief states of the stereotype members can be recovered. For example, if 6 out of 10 agents believe grass is green, the distribution for 10 agents given a mean value of 0.6 is 6 out of 10. Happily, the mean value can be used directly by the evaluator, as the belief probability value, to find the expected utility over the stereotype members for a game tree.
A variation of using the mean value is to use a decaying average of evidence from recent dialogues. For example, each revision of the stereotype might use a weighted sum of ninety-five parts of the previous stereotype value, with five parts of the belief state produced by the current dialogue.
The system normally uses implicit acquisition, where it passively observes the dialogues and obtains information from belief revision. There is no use of explicit acquisition dialogue, such as direct questions, although such questions and the value of information they provide could be easily computed within the system, just by writing some plan rules for the explicit questions that occur when a new user is introduced, and evaluating the game tree that results. Explicit acquisition is discussed more in the "future work" chapter.