In previous experiments  evolution of variable length mathematical expressions containing input variables was found to be useful in finding learning rules for simple and hard supervised tasks. However, hard learning problems required special attention in terms of their need for larger size codings of the potential solutions and their ability of generalisation over the testing set. This paper describes new experiments aiming to find better solutions to these issues. Rather than evolution a hill climbing strategy with an incremental coding of potential solutions is used in discovering learning rules for the three Monks' problems. It is found that with this strategy larger solutions can easily be coded for. Although a better performance is achieved in training for the hard learning problems, the ability of the generalisation over the testing cases is observed to be poor.
Download compressed postscript file