In this study, applying the Varaiya's decomposition principle to the grid networks made up of many closed loops, the offset patterns of the traffic signals are optimized. The total delays of the networks are minimized subject to the closure constraints. It is supposed that the delays of links are the functions of only the relative offsets, and that the delays are approximated by two parabolic curves. The networks are decomposed into the many subareas of single loops. At the lower level, these single loops are optimized independently each other, and the coordination between the subareas is performed at the upper level The offset patterns are obtained by repeating the two level computations. From some examples, it is found that the convergences to the solutions are achieved within reasonable CPU time.