The primary objective of this research has been to discredit the Occam thesis. To this end it uses a post-processor that disregards the Occam thesis and instead is theoretically founded upon the similarity assumption. Experimentation with this post-processor has demonstrated that it is possible to develop systematic procedures that, for a range of `real-world' learning tasks increase the predictive accuracy of inferred decision trees as a result of changes that substantially increase their complexity without altering their performance upon the training data.
It is, in general, difficult to attack the Occam thesis due to the absence of a widely agreed formulation thereof. However, it is far from apparent how the Occam thesis might be recast to both accommodate these experimental results and provide a practical learning bias.