In this thesis, we study channel estimation in time-varying cooperative network. Since channels vary with time, we insert training blocks periodically to trace channel variation. In this work, we adopt Kalman filter to trace channel variation due to its low complexity. By storing previous channel estimate, Kalman filter simply requires to process next received vectors to update current channel estimate. We use all past observations to estimate current channel state to avoid wasting information. In content of cooperation, we directly estimate effective channel from source through relay to the destination. The reason is that, we separately estimate the source-relay and relay-destination links, relays need extra efforts to estimate the channel and feedback estimates to the destination. It will increase the computational loading on relays, and the feedback channel may suffer channel fading, resulting in more distortion of estimates. Therefore, the destination directly estimate effective channel, using Kalman filter to trace variation. Furthermore, we design pre-coding scheme on relays for forwarding training symbols in order to reduce channel estimation errors and obtain more accurate channel information. To detect data symbols, we need to channel state information over each data block as well. Therefore, estimates over previous training blocks are interpolated to estimate channel over data blocks based on LMMSE criterion. Since estimates over training blocks are obtained from Kalman filter, it consequently improves estimation quality of the channel over the data blocks. The main contributions of the thesis are optimal training design to reduce the estimation error, the estimation based on Kalman filter, and linearly combing the estimates to provide more accurate estimates of the channels over data blocks.