M

Abstract

A supervised learning algorithm which uses an automatic inductive approach to recognise patterns in data (Cracknell and Reading 2014); it is a means of prediction based on the properties learnt from a large set of training data, and performance is judged by its ability to reproduce known knowledge. Ideally, such a system should be able to improve its performance over time, learning from its mistakes. Although the term “machine learning” was introduced in the late 1950s (Samuel 1959; Felsen 1976), it was the development of the Boltzmann machine (presumably named for the Austrian physicist, Ludwig Eduard Boltzmann (1844–1906), whose development of statistical mechanics showed how the properties of atoms determine the physical properties of matter), a type of artificial neural network and Markov random field network consisting of symmetrically connected, neuron-like units that make stochastic decisions about whether to be on or off. It has a simple learning algorithm which allows it to discover interesting features which represent complex regularities in the pattern vectors of the training set. It also contains “hidden” units which are not directly influenced by the values of a pattern vector but rather by the values assigned to the “visible” nodes which are so influenced (Hinton and Sejnowski 1983; Ackley et al. 1985; Michalski et al. 1983). This work led to the present generation of algorithms which have extended the approach. Earth science applications are discussed by Oommen et al. (2008), Waske and Braun (2009), Leverington (2010), Leverington and Moon (2012), Yu et al. (2012) and Cracknell and Reading (2014). See also naïve Bayes, k -nearest neighbours, random forests and support vector machines, pattern analysis, pattern classification, pattern recognition.