A direct algorithm to determine the weights of multi-layer feedfoward neural networks based on logical input samples is proposed in this paper. It can be realized by only three-layer networks for logical input samples; otherwise

a pretreatment layer between input layer and hidden layer has to be introduced to transform input samples into logical vectors. Because error-energy function is not introduced

the problems of slowly convergent rate and error trapped in to local minimum appearing in BP algorithm are avoided in this algorithm. Finally

the correctness of the algorithm is proved in detail and the computing process is explained with XOR problem.