Sistemi di Apprendimento Simbolico-Neurali: Reti Neurali per il Ragionamento Normativo

In real life scenarios relative to normative contexts, norms are not always explicitly described. This often occurs in the legislative field, where judges takes into account tacit norms to make their trials. In fact tacit norms are not explicitly described in the nomative code used to formulate the trial. Due to the lack of this tacit knowledge in normative codes, judges trials may result as made without truly following the normative regulations contained there. This thesis aims to address this problem using the Neural-Symbolic Integration, this approach allows to take advantage of both symbolic logic and connectionist formalisms. The symbolic approach can be used to represent the normative code that contains the explicit norms followed by the judges, then exploiting the work made by d'Avila Gacez, Gabbay and Broda, we can translate the symbolic representation of the normative code in a neural network able to compute the same semantic. In this way we can exploit neural networks capacity to learn from instances and train them using exsting trials, the trained network, using the starting and the aquired knowledge will be able to produce trials on future cases as a judge will do. In addition, from the trained network, again following the work of d'Avila Gacez, Gabbay and Broda, will be possible to extract the improved knowledge contained and build a new normative code that can be used for a more effective normative reasoning using other formalisms and to explicitate tacit norms learned by the neural network. The work contained in this thesis will also take into account contrary to duty problems and dilemmas relative to Deontic logic (and its extensions) used to model obligations and permissions, the core of normative reasoning. Considering this problems, we will propose a method using a priority-based ordering between the rules able to handle them both using the symbolic representation and in the neural networks obtained from the translation.