Energy used by air conditioning and especially cooling is steadily increasing, and modernization of the technical systems is critical. The standard controllers must be replaced by novel and more efficient controllers. In the present article a model-predictive controller (MPC) called NeuroCool is presented, which features a self-learning building thermal model. In this article an overview of the algorithm and associated simulation / testing environment is provided; simulation results are analyzed to assess the performance of the algorithm. It is shown that when benchmarked against a standard controller, exploitation costs can be reduced by about 11% under similar comfort. If the comfort is lowered, but maintained within the tolerable norms, exploitation costs can be further reduced by 75%.