Previous research efforts, for optimizing energy usage of HVAC systems, require either mathematical models of HVAC systems to be built or they require substantial historical operational data for learning optimal operational settings. We introduce a model-free control policy that begins learning optimal settings with no prior historical data and optimizes HVAC operations. The control policy is an adaptive hybrid metaheuristic that uses real-time data, stored in building automation systems (e.g., gas/electricity consumption, weather, and occupancy). It finds optimal setpoints at the building level and controls set points accordingly. The algorithm consists of metaheuristic (k-nearest neighbor stochastic hill climbing), machine learning (regression decision tree), and self-tuning (recursive brute-force search) components. The control policy uses smart selection of daily setpoints as its control basis, making the control schema complementary to legacy building management systems. To evaluate our approach, we used the DOE reference small office building in all U.S. climate zones and simulated different control policies using EnergyPlus. The proposed algorithm resulted in 31.17% energy savings compared to the baseline operations (22.5 C and 3 K). The algorithm has a superior performance in all climate zones for the goodness of measure (i.e., normalized root mean square error) with a value of 0.047. (C) 2017 Elsevier B.V. All rights reserved.