Exploring the connections between symbolic and subsymbolic reasoning in producing data-driven decision making.

The core motive is investigation of the potential advantages of a hybrid reasoner that can draw on the power of machine learning techniques whilst still benefiting from relatively human-friendly logic. Principally the research has begun with attention very much focused on argumentation as the symbolic layer due to its accord with the properties of defeasible knowledge. By stripping away assumptions about inherent logical structures and instead focusing on the nature of attacks within arguments a simple model is established that can harmonise with complex algorithms without losing the most fundamental property of knowledge - that of truth and falsehood.

At this stage the research is thus primarily aimed at exploring machine learning theory in light of potential connections with argumentation structures. A neural network architecture has been proposed that captures the formulation of argumentation as per Dung semantics; attacks are represented by edges and can be learned from argument acceptability data. This implementation will be extended to more complex forms of argumentation that more accurately reflect the defeasible nature of knowledge by permitting doubt to be expressed with respect to the values associated with arguments and attacks as well as variability in their individual strengths. Further research on advanced argumentation techniques is required with particular attention given to numerical based argumentation.

Finally the research will look to develop the theory into a viable application in a domain that permits comparison with extant methods.