Evolutionary algorithms and neural networks. Theory and applications (Q721117)

Evolutionary algorithms become a popular tool in finding optimal solutions for optimisation problems. They represent an alternative to the well-known Backpropagation algorithm. The evolutionary algorithms allow for the determination of a good architecture, setting the number of units and the number of hidden layers in the feed-forward networks. The search space may contain many sub-optimal solutions corresponding to local minima that the evolutionary algorithms may avoid. The book is composed of two parts. In the first part, evolutionary algorithms for optimisation are presented, and in the second part they are used to train neural networks. The nine chapters have the structure of scientific papers with a corresponding bibliography at the end. In the first chapter a general introduction to the evolutionary single-objective algorithms is presented where a global optimum for single-objective problems is searched. Constraints define which of the values for the input are valid. The difference between individual-based and population-based is highlighted. Exploration and exploitation are in conflict, a good balance in an algorithm is required. The search is done locally around the most promising solutions found in the exploration phase. Abrupt changes of the candidate solutions ensure that they explore different regions of space. In the following four chapters four popular evolutionary algorithms are presented. Particle swarm optimisation simulates the navigation of birds in flying swarms. Particles are allowed to move in a search landscape by the principles of separation, alignment and cohesion. In the next chapter the ant colony optimisation is introduced for combinatorial optimisation problems. The algorithm is based on stigmergy. Stigmergy refers to the manipulation of the environment by biological organisms to communicate. Routes of ants from a nest to a food source are marked by pheromones. The pheromone decrease is proportional to the length of the path. In chapter four the basic principles of genetic algorithms are introduced. The entire population gets better generation by generation. The algorithm is based on the crossover operation that ensures a local search and mutation that leads to a global search. The selection of each generation is done according to its fitness value: higher fitness value corresponds to the probability to be selected. The last one of the evolutionary algorithms in chapter five is the biogeography-based optimisation that is as well based on two operators that perform crossover and exploitation. However, it takes additionally into account the geographical distribution by immigration and emigration and mutation to reach an optimal balance between preys and predators in an ecosystem. In the second part the four presented algorithms are applied to train neural networks. In feed-forward networks, instead of Backpropagation, evolutionary algorithms are used to determine the weights and biases. It is an alternative that may avoid local minima. In the seventh chapter six function approximation datasets are tested. Then, in the following chapter the evolutionary Radial Basis Function networks are described. The last chapter describes experiments with evolutionary deep neural networks on a hand-gesture detection dataset. The evolutionary algorithm performs the feature selection and determines a good number of layers. The book is a short and clear introduction to the domain of evolutionary algorithms and their application to train neural networks.