The performance of a pneumatic servo-system is affected by compressibility of air, friction between cylinder and piston. Due to these non-linear factors, it would be difficult to accomplish the satisfactory control performance by using PID or adaptive control theory. Fuzzy control is one of effective control theory for such non-linear plant. However, synthesis of fuzzy rule including stability has not been well established yet. On the contrary, Neural network is excellent in learning performance. In this paper, we consider a learning method to acquire the appropriate fuzzy rules using error back propagation to improve control performance of pneumatic servo system. In the method, two criterions are defined and the fuzzy rules are adjusted so as to minimize them respectively using error back propagation. Moreover, differentiating coefficient of the plant used in error back propagation is acquired by newly established neural network. The method is applied to the vertical type pneumatic servo systems to prove it's effectiveness.