Combining the Advantages of Neural Networks and Decision Trees for Regression Problems in a Steel Temperature Prediction System

Simple decision trees enable obtaining simple logical rules with a limited accuracy in regression tasks. Neural networks as highly non-linear systems can map much more complex shapes and thus can obtain higher prediction accuracy in regression problems, that is however at the cost of the poor comprehensibility of the decision process. We present a hybrid system which incorporates the features of both a regression tree and a neural network. This system allowed for achieving high prediction accuracy supported by comprehensive logical rules for the practical problem of temperature prediction in electric arc furnace at one of the steelwors.