Master Thesis Defense by Mr Luis Palacios Medinacelli
Mr Luis Palacios Medinacelli defended his master thesis on ' Skeptical Abduction: A Connectionist Network'.
Mr Luis Palacios Medinacelli defended his master thesis on ' Skeptical Abduction: A Connectionist Network' at TUD on 23 March 2016.
Abstract: A key research area in Artificial Intelligence is human thought. Providing theoretical models for human reasoning is of great interest in AI as it sheds light on the ongoing processes in our minds and brains. Two fields studying these processes are Machine Learning and Knowledge Representation, whose approaches are usually incompatible. Thus a model that can take advantage of both fields is desirable. In this thesis we provide a Neural-Symbolic realization for computing Skeptical Abduction. The resulting network is a prerequisite to model some human reasoning tasks presented by Byrne. Some of the afore mentioned human reasoning tasks were modeled in a cognitive-computational approach by Stenning and van Lambalgen. In particular they show that Byrne’s tasks can be adequately modeled by logic programming. Their approach was modified by Hölldobler and Kencana Ramli using ?ukasiewicz three-valued logic and weak completion semantics, and extended to a Neural-Symbolic representation. Their approach covered most of the tasks, but not all, since some of them require skeptical abduction. This thesis solves the problem of how to obtain a Neural-Symbolic system for computing skeptical abduction, and thus completing the work by the above mentioned previous approaches, which constitutes our main contribution. Our approach uses abductive logic programming to formally specify skeptical abduction. Then we provide an algorithm to compute skeptical inferences, that is encoded as a logic program. Once the problem is solved at a logic level, we use the Core method to translate the resulting logic programs into a neural network. Both, the algorithm and the construction of the network heavily rely on the properties of semantic operators for logic programs. As a result we obtain a neural network that adequately models Byrne’s human reasoning tasks that require abduction. The network can also be used for solving additional problems related to abduction. It can as well be trained, meaning that the knowledge represented by the network can be learned.