Combining genetic algorithms and neural networks : the encoding problem

Neural networks and genetic algorithms demonstrate powerful problem solving ability. They are based on quite simple principles, but take advantage of their mathematical nature: non-linear iteration.

Neural networks with backpropagation learning showed results by searching for various kinds of functions. However, the choice of the basic parameter (network topology, learning rate, initial weights) often already determines the success of the training process. The selection of these parame- ter follow in practical use rules of thumb, but their value is at most arguable.

Genetic algorithms are global search methods, that are based on principles like selection, crossover and mutation. This thesis examines how genetic algorithms can be used to optimize the network topology etc. of neural net- works. It investigates, how various encoding strategies influence the GA/NN synergy. They are evaluated according to their performance on academic and practical problems of different complexity.

A research tool has been implemented, using the programming language C++. Its basic properties are described.