A story of the artificial ant: discovering the correct bias for learning

The artificial ant problem is one of the benchmark problems where an artificial ant is evolved to learn to collect a set of food pieces placed along a particular trail. The degree of generalisation of the learned behaviours of the ant to similar trails has been an important issue of concern among several researchers. We present series of experiments and analyses them in terms of two well established machine learning concepts: generalisation and learning bias. The issue of generalisation is directly related to learning bias. Without a proper bias there is a great risk that, for a given problem, generalisation may not be possible. The experimental results show that finding a correct bias in evolutionary experiments may improve generalisation of genetic based simulated learning behaviours.