Both simulation and expert systems are popular ways to solve complex and hard problems. However, the results of the simulation, which include a large amount of valuable information as a good knowledge source, are not used efficiently. Furthermore, the development of the expert systems can fail because there is no expert or an expert is not available. A new Simulation-Based Expert System(SIMBES) paradigm has been constructed to overcome these problems. It consists of simulator, feature extractor, machine learning system, performance evaluator and Knowledge-Based Expert System(KBES). A SIMBES was implemented for an existing schedule-based MRP system in Smalltalk/V to show how this paradigm works and experimented for a large number of jobs. The KBES and the existing system produced better schedules for 72 percent and 28 percent of the jobs, respectively.