The goal of this Ph.D. work is to improve multiple criteria decision making by the use of " context-dependent " preference models. Such models are more realistic than the ones used before, but it is usually difficult for the decider to express them. This is the reason for using machine learning by neural networks for identifying the preference model. A neural net " learns " the preference model by the principles of non-linear regression, where the decision model is expressed by examples of decisions. The INKA neural network developed reduces the number of examples necessary, which is essential for the practical use of this technique. Learning times are also sufficiently short to make possible the interactive acquisition of the preference model. The interactive decision support system developed using INKA is one of the first to use machine learning to identify a global preference model. The visualisation of the learnt model and the indicators of precision and sensibility developed help the decider to decide when to stop the interactive procedure. This is especially important for learning the preferences of " abstract deciders " (a group of people, consumers, nature, ...), who can not interact with the system. Explaining the results is still a great problem both for decision support systems and for neural networks. The methods developed here make it possible to reduce or eliminate this problem. It is therefore possible to explain, understand and analyse even the preferences of " abstract deciders ". This information may then be used for improving group decision making or for improving product sales which depend on consumer preferences.
