Abstract
Abstract
We show the applicability of neural networks for distance estimation in classical search
problems. First, we present and evaluate different techniques which are able to sample
training data from difficult problems of arbitrary domains. Afterwards, an empirical
investigation on good neural network configurations for learning a goal dependent heuristic
is performed. Finally, the trained networks are evaluated as heuristics in actual searches
and compared to state of the art techniques. We have observed that for difficult problems
the neural networks perform faster searches and generate better plans than other state of the art techniques.
