A modification to the geometrical learning algorithm of binary neural network

which tries to enhance efficiency of the algorithm

is demonstrated. Then a new Heuristic Genetic Geometrical Learning algorithm（called HGGL algorithm） of the neural network used for arbitrary Boolean function approximation is presented. The algorithm imtroducesknowledge based crossover operator and mutation operator to optimally divede geometrical hypercube and evaluate the numberof the hidden netirons

connection weight and threshold. For arbitrary Boolean function

the neural network trained by HGGLalgorithm has the fewest number of hidden layer neurons comparde with existed leaning algorithms.