This study investigates the optimal setting of green times for traffic lights in an isolated intersection with the purpose of minimizing congestion. A machine learning method forms the backbone of the proposed method. Here, Q-learning is applied for signal light timing to minimize total delay. It is assumed that an intersection behaves similar to an intelligent agent learning to plan green times in each cycle using current traffic information. Compared to previous studies in this field, we expand the state space and innovatively set the reward to the average difference between traffic that enters the intersection and the queue length in the corresponding links. In contrast to previous studies, it is also assumed that the cycle time is variable. The performance of the proposed method is comprehensively compared with two traditional alternatives for controlling traffic lights. Simulation results indicate that the proposed method significantly reduces the total delay in the network when compared to the alternative methods.