This master thesis surveys multilayer perceptrons and associated training algorithms. Factors influencing generalization and convergence properties are discussed throughout the thesis. Selected heuristics are described in detail outlining conditions that effect their success. Dedicated sections demonstrate some approaches to unify stochastic algorithms - like Simulated Annealing and Genetic Algorithms - and neural networks. Emphasis is directed at the interaction and dependencies between network components and training parameters. In a practical section neural networks are used to introduce a mapping between the chemical composition and physical characteristics of aluminum alloys. The goal is to find a network structure that allows accurate prediction of mechanical and casting properties. A variety of algorithms introduced in preceding chapters are trained and evaluated on real-world training data. Conclusions about convergence behavior as well as generalizations are drawn.