Solving Constraints with Trigonometric Functions Occurring in the Workspace of a Mobile Robot by Methods of Machine Learning

The paper represents first results on solving constraint nets consisting of equations and inequations containing trigonometric functions by methods of Machine Learning. Constraints of this type occur for example in planning movements of a mobile robot on a symbolic level in a workspace where obstacles or other “dangerous regions” have to be avoided. Another application area is the automatic generation of layouts and diagrams from textual input. Learning proceeds by “active” exploration of the workspace by generating training data classified by constraint satisfaction vs. non satisfaction. Results obtained with both decision tree learning and a neuronal net of the Perceptron type demonstrate good approximation and generalization properties depending on the number of constraints and training data.