Due to the existence of the uncertainty of the traffic flows in the problem of elevator group control scheduling, markov decision processes cant deal with the model of the elevator group control scheduling well. The model based on partially observable markov decision processes is studied to solve this problem. With application of the feedforward neural network,which is integrated into the linear Q-learning to construct the whole algorithm for elevator group scheduling, the optimal or nearly optimal policies are obtained gradually. Compared with previous algorithms, this algorithm greatly improves the adaptability to the up-peak traffic flow.