Deep reinforcement learning (DRL) has become a popular optimal control method in recent years. This is mainly because DRL has the potential to solve the optimal control problems with complex process dynamics, such as the optimal control for heating, ventilation, and air-conditioning (HVAC) systems. However, DRL control for HVAC systems has not been well studied. There is limited research on the real-life implementation and evaluation of this method. This study implements and deploys a DRL control method for a radiant heating system in a real-life office building for energy efficiency. A physics-based model for the heating system is first created and then calibrated using the measured building operation data. After that, the model is used as a simulator to train the DRL agent. The trained agent is then deployed in the actual heating system, and a smartphone App is used to let the occupants submit their thermal preferences to the DRL agent. It is found the DRL control method can save 16.6% to 18.2% heating demand compared to the old rule-based control logic over the three-month deployment period. However, several limitations of this study are found, such as the low participation rate of the App-based thermal preference feedback system, inefficient DRL training, and the requirement for a large amount of building data.