A hybrid neuro-genetic pattern evolution system applied to musical composition

The two processing paradigms of Genetic Algorithms (GAs) and Neural Networks (NNs) are combined to form a hybrid system to act as an adaptive pattern evolution device; this system is subsequently applied to a musical composition task. GAs are well suited to algorithmic musical composition, due to the complex nature of the search space of this task. Previous applications of GAs to musical composition have produced satisfactory, yet limited results. The limitations of these results are caused by the fitness functions implemented by genetic composition systems. The types of fitness function used in musical GA applications are identified and analysed here, along with reasons for the limitations of each. A novel means of evaluating the fitness of candidate solutions in a evolutionary musical application is presented the Adaptive Resonance Theory neural network. The ART neural network assigns fitness to individuals in a population according to the degree of similarity displayed between the individual and information used to train the network. Additionally, the ART neural network paradigm is adaptive to novel classification types, with no degradation of existing classification knowledge. Comparisons with the Multi-Layer Perceptron (MLP) neural network, common in other neuro-genetic systems, confirm the advantages of using ART networks as fitness evaluators: ease of training (and re-training), no minimum amount of training data, few network-specific parameters, and fitness measurement relevant to the search process. Analysis of this fitness evaluation process reveals that the ART neural network performs a niching process at the genotypic level, in contrast to other genetic search tasks, where niching is carried out at the phenotypic level. The application of this system to a musical composition task is demonstrated; this application generates realistic multi-voiced percussion sequences, using existing percussion sequence information from commercially available drum machines as training data for the fitness evaluator.