Semiconductor manufacturing fab is one of the most sophisticated man-made system, consisting of hundreds of very expensive equipment connected by highly automated material handling system. Operation schedule has huge impact on the productivity of the fab. Obtaining efficient schedule for numerous equipment is a very complex problem, which cannot be solved by conventional optimization techniques. Hence, heuristic dispatching rules combined with fab simulation is often used for generating fab operation schedule. In this paper, we formulate the fab scheduling problem as a semi-Markov decision process and propose a reinforcement learning method used in conjunction with the fab simulator to obtain the (near-)optimal dispatching policy. Resulting schedule obtained by the proposed method shows better performance than heuristic rules whose parameters are tuned by human experts.