Structural Generalisation with Decision Forests
William Uther

Decision trees have a problem known as 'small disjuncts'.  A decision tree
works by recursively splitting the supplied dataset.  If there is a sub-concept
in the concept being learnt that is present in different parts of example space
it has to be learnt multiple times - once for each area of feature space the
sub-concept is in.  This happens frequently in boolean formulae for example. 
Moreover, each instance of this sub-concept is being learnt with only partial
data, the data that fell in that sub-space, rather than all the data describing
that sub-concept.  I'll talk about a method by Oliver (1992) which helps
alleviate this problem by inferring decision graphs using the Minimum Message
Length principle.  For Reinforcement Learning something more is needed though. 
I'll then discuss my method for inferring decision forests - a more general
representation than decision graphs.  My work is still in its early stages and
I expect the chat to be fairly informal...