Music generation using deep learning technology had become an important application in Artificial Intelligence. In order to have reasonable pleasant melody, basic music theory could be incorporated in any computer algorithms. One important issue is the generation of melodic skeleton, which could be an important portion of the original melody. We propose to use an optimization technique, names simulated annealing, to compute melodic skeleton based on a given melody. In order to justify the feasibility, we developed a web-based tool ( http://140.115.51.218:8889/ ), which allow users to generate two types of Taiwan’s pop music. The website uses deep learning technology to generate main melody, followed by the algorithm proposed in this article, to generate Heterophony Accompany as the second melody. Additional features such as functional harmonic and percussion instruments are also included. We also evaluate the outcome of the simulated annealing algorithm, with two quantitative methods. The outcome justifies our feasibility. Most importantly, the developed tool was used by many musicians to generate practical cases used in the industry.