Figure 10:
Plot Showing Population Means For Test Data
In this work we have applied a Pareto-based MOGA to parameter selection for the PGH object recognition paradigm. This is not a trivial problem as the set of optimal histograms is defined recursively by all others in the database. Heuristics were applied in the crossover and mutation operators to avoid speciation and the need for fitness sharing. By carefully selecting our data sets we have shown that the optimisation technique works irrespective of many common problems such as noise, missing data and segmentation. The effect of applying the MOGA is the optimisation of a set of pairwise parameters which until now could only be specified by the user. We are therefore making the construction of PGHs a fully automatic, optimised process. The algorithm produces distinct, efficient histogram definitions for subsets of lines in our database. Potentially it could be used to learn to recognise objects in an environment in which classification feedback is only available for a cluttered (un-segmented) scene (ie: a real-world learning environment).
Frank Aherne