Validation of agent-based models

Abstract

The automatic collection of customer transaction data, through either online shops or reward cards, is producing very large databases which contain much information about consumer behaviour. What kind of information and how exploitable it is are very relevant questions. Two approaches are being used. Either one concentrates on individual behaviour and tries to apply various theoretical frameworks and results of the literature on discrete choice, or one uses clustering algorithms in order to determine several classes of customers. The very existence of such categories is likely to be the result of social interactions and influences. The literature on discrete choice cannot easily be generalised to networked interactions, which are known to be widely present in various contexts. Another approach is to use toy models of individual behaviour and concentrate on global, aggregate quantities such as market share or demand fluctuations. This raises the question of how to validate such kind of model, hence the request of Unilever. The latter should also be understood with respect to the contribution of ESGI 2004, where a very sophisticated agent-based model of consumer behaviour was proposed (but not much studied).