Under the fast fading environment, the estimated channel state information (CSI) is largely different from real channel state particularly in the last part of the packet. To mitigate this influence, we previously proposed a multilayer feedforward neural network (MLFNN) based channel estimation method. Regression capability of the MLFNN well estimated the whole transition of CSI. This network is trained by using a few CSI data set at beginning part of the packet. These partial CSIs are obtained by the pilot-aided channel estimation (PCE) and the decision feedback channel estimation (DFCE). However, MLFNN back-propagation (BP) training needs iterative renewal process of parameters. Thus, the computational complexity of the training part is quite large. To overcome this problem, this paper newly proposes a generalized regression neural network (GRNN) based channel estimation for OFDM system. Because of the direct detection method for parameters applied to GRNN, it can estimate the whole transition of channel states without huge complexity training and the processing delay. The computer simulation results clarifies that the proposed method can improve the BER performance even while the calculation quantity is minimized.