The thesis has presented a thorough research of Deep Reinforcement Learning used in conjunction with a Digital Twin for the optimization of combustion efficiency and emission control in a gasoline engine. The study was inspired by the limitations of conventional map-based calibration and the need for adaptive, intelligent control systems to meet future regulations and real-world performance requirements. A Soft Actor-Critic agent was efficiently trained to control the spark timing and air-fuel ratio through a well-designed experimental plan. The policy learned by the model showed that it had several levels of understanding of the physical concepts by following the lead of the "retard of the spark" and "lean-burn" operation to get the result that is a good compromise between efficiency and emissions. The comparisons on the basis of the conventional controller have shown that the Deep Reinforcement Learning controller is way ahead in performance and especially on average NOx emissions have been reduced by 34.7% in different driving cycles without yet a statistically significant fuel economy penalty from the same experiments, and even the efficiency in aggressive driving has been improved. Most importantly of developing and demonstrating the practicality of deep reinforcement learning voice of control in car combustion engines. The hardware-in-the-loop validation demonstrated the online feasibility of the RFC being a thousand times faster than the requirement with a small memory footprint suitable for production ECUs. In summary, the thesis not only opens a new window into how to combine and harness the technologies within a closed loop engine control system but also convincingly argues for considering machine learning as the preferred method for the control of the next generation of engines. By taking into account the instantaneous nature of the unguided, shared and overall world's driving conditions, it provides a recipe for the synthesis of high efficiency, low emissions, and dynamic performance modes, which are often in conflict. The main obstacles lying in the path of the widespread deployment of DLC are certification and full physical validation, but the work here already goes a long way to achieve that if not fully.