Direct torque control (DTC) is the simplest torque control of induction machines forindustrial application. The key component of DTC is the state selector. In this paper, wewill discuss to use the neural network to emulate the state selector of the conventionalDTC, and train the neural network with a genetic algorithm. In the genetic algorithm, weused floating point encoding and some specialized operators (e.g., non-uniformmutation, arithmetical crossover, etc.), and designed a new operator - non-uniformarithmetical mutation. It greatly improved the fine local tuning capabilities of a geneticalgorithm. The experiments of simulations have been carried out in the same machinewith the conventional DTC and the trained state selector neural network. The simulationresults show that the results of this paper is far better the outcome given by