Questions: Are artificial neural networks useful for the automatic assignment of species composition records from vegetation plots to a priori established classes (vegetation units)? Is the assignment more accurate (1) if the classes are defined by numerical classification rather than by expert-based classification; (2) if the training data set is selected to include plots that are richer in diagnostic species of particular classes? Material: Species composition records (relevés) from 4186 plots of Czech grasslands. Methods: Plots were classified into 11 phytosociological alliances (expert classification) and into 11 clusters derived from numerical cluster analysis. Some plots were used for training the classifiers, which were the multi-layer perceptrons (MLP; a type of artificial neural network). Other plots were used for testing the performance of these classifiers. Plots used for training were selected (1) randomly; (2) according to higher representation of diagnostic species of particular classes. Results: Different MLP classifiers correctly classified 77-83% of plots to the classes of the expert classification and 70-78% to the classes of the numerical classification. The better result in the former case was mainly due to two classes in the expert classification, which were well recognized by the classifiers and at the same time contained a large proportion of the plots of the entire data set. Correct classification of the plots belonging to these large classes resulted in a good overall performance of the classifiers. After training with randomly chosen plots, the classifiers produced better results than after training with plots that contained more diagnostic species. This indicates that the biased selection of the training plots disables the classifiers to recognize the entire variation within the classes and results in errors when new plots are to be classified. Conclusions: MLP is suitable for assigning vegetation plots to already established classes. Unlike some other methods of supervised classification, it performs well even in communities that are poor in diagnostic species. However, the method does not provide clear assignment keys that could be used for class identification in field surveys. It is therefore more appropriate in applications that aim at a reliable class assignment rather than understanding the assignment rules.