Thesis Proposal Update Communication of Inductive Inference

The original proposal described the goal of the thesis as an investigation into means to combine local inductive hypotheses into a single globally consistent hypothesis. The reviewer agreed with the defined problem, and furthermore supports the goal. The proposal advocated the use of version space intersection as the integration method. Version space intersection is the theoretically correct approach, but as with many theoretically correct approaches in AI, it has turned out that for practical usage that it is too computationally expensive as it relies on a near exhaustive search to deal with all all possible scenarios. More recent and extensive experimental results demonstrate this, and has exposed another more fundamental problem: On several well known machine learning data sets (mushrooms & congressional voting records) the version space intersection approach was not only computationally expensive, but more importantly, was unable to achieve useful summarization of the training data. In this we have reached experimental agreement with the reviewer. We also agree with her that in that certain domains, where the hypothesis language is restricted, that such an approach might still be useful. However, whilst Hirsh developed the version space intersection approach and in later work showed that it was not very practical, he never discussed its application to a distributed agent setting, which is why we carried out our initial work. We now have experimental evidence that it is not an effective method to deal with the problem of integration of distributed hypotheses (which may be why Hirsh did not consider it, but this is conjecture).