ABSTRACT

Confirmation, Resolution and Decomposition: Handling Inconsistency with Argumentation

Robert Bowerman Computer Science UCL

Many problems in artificial intelligence (AI) start with a knowledge base of logical formulae, from which conclusions are deduced. Deduction is one of the key ways that, at times, can make computers appear to be intelligent. However, two difficulties commonly occur with this approach: the first is inconsistent deductions and the second is repetitions in the deductions. Imagine you are reading a newspaper that keeps repeating itself and contradicting itself -- you would not have much confidence in what you read! Although such repetitions could be simply deleted they tend to contain valuable information, information that can be used for resolving inconsistencies.

A solution approach is to structure the deductions as arguments, whereby each distinct argument has a set of premises and a conclusion. Argumentation is emerging as an important part of AI for working with uncertain knowledge. It is common to find a variety of arguments for and against a given proposition. How to accumulate these competing arguments to draw a firm conclusion is an area of much research focus in AI. Argumentation is emerging as a strong technique for handling inconsistency and uncertainty.

This talk looks at a) a way of representing arguments and their aggregations, b) various naove argument aggregation schemes from the literature and c) some novel proposals for more effective, true-to-life ways of aggregating arguments. Motivating examples are given from various professions where the use of argument is widespread. Of particular interest is using arguments that confirm each other to establish that a confirmed proposition is more definite than an unconfirmed one. These aggregations are then used to resolve inconsistencies, in a way akin to a judge deciding which side has more solid evidence. The other novel approach to be examined is decomposition, where an inconsistency arises from the muddling of two contexts.


Maintained by rbennett@cs.ucl.ac.uk