Simon Parsons; Natalia Criado; Maite López-Sánchez; Juan A. Rodríguez-Aguilar

Currently, the Internet and its virtual platforms are the main form of communication in our lives. From international to local communities, citizens search and demand for better ways to express their opinion and to decide about the world they live in, to decide collectively. However, current collective decision making methods have yet to improve to achieve their potential.

Inspired by e-participation systems, this dissertation explores multi-agent debates and collective reasoning on them. Three novel approaches to represent a multi-agent debate ---the Target oriented discussion framework, the Relational model and the Abstract multi-agent debate---  are presented and used to study collective reasoning methods. The use of dependencies within a debate and coherence, a notion to capture opinion consistency, play a key role throughout this research. 

The Target oriented discussion framework structures an argumentation-based debate allowing both positive and negative relationships between the arguments and making it possible for participants to express their opinions about the arguments. In particular, it  addresses the problem of how participants can reach agreement about a single issue being discussed. Several new methods to reach a collective decision are assessed by means of social choice properties. Further to the analysis, a computational assessment shows their applicability in real scenarios.

The Relational model overcomes drawbacks of existing approaches, by leaving aside arguments and attack and defence notions, to arrange a more general representation of a multi-agent debate. This model provides a clear distinction between different features composing a debate while offering more expressiveness to participants. A family of new opinion aggregation functions is defined and an exhaustive analysis of their performance regarding their social choice properties is provided. Concluding, a computational  analysis demonstrates that collective opinions can be computed efficiently for real-sized debates.

Finally, the Abstract multi-agent debate model extends the notion of a multi-agent debate allowing it to be an abstraction for different approaches. After proving its capability to represent other debate models, an approach to analyse the quality of a multi-agent debate is introduced.