In this thesis, we try to model the problem of constructing a multicast tree with minimum cost. Unlike the other minimum cost multicast tree algorithms, this model consists of one multicast group of fixed members. Each destination member d is dynamic and has a probability of being active as d q which was gathered by observation over some period of time. With the omission of node join/leave handling, this model is suitable for prediction and planning purpose than for online maintenance of multicast trees. Lagrangean relaxation method is applied to speed up the solution procedure.