Fast Evolutionary Learning with Batch-Type Self-Organizing Maps

摘要

Although no distance function over the input data is definable, it is still211u001epossible to compute the self-organizing map (SOM) using evolutionary-learning 211u001eoperations. The number of computations required is decreased, and the 211u001eevolutionary process is made to converge rapidly when the probabilistic trials of 211u001econventional evolutionary learning are replaced by averaging using the so-called 211u001eBatch Map version of the Self-Organizing Map. An order in the map that complies 211u001ewith the 'functional similarity' of the models can be seen to emerge. The spatial 211u001eorder in the array of models can be utilized for finding more uniform variations, 211u001esuch as crossings between functionally similar models. There exist two modes of 211u001euse of this new principle: representation of nonmetric input data by models that 211u001emay have variable structures, and very fast generation of evolutionary cycles 211u001ethat resemble those defined by the genetic algorithms.