Abstract: Aiming at the curse of dimensionality caused by prodigiousness of state-action space for Markov decision-making process model, a kind of Q learning method based on neural network was proposed. The Q value of a state-action pair during reinforcement learning was approached and stored by means of the high generalizability of BP neural network, then the optimal strategy based on Q learning for selection of action and a BP neural network model and algorithm for Q learning were designed. The algorithm proposed was applied to task allocation of process planning, with a simulation done by the software Matlab. The result indicated that it has a good performance and the capability of action approach, and the method enhances the applicability of reinforcement learning in task allocation.