Effectively controlling supermarket refrigeration systems is critical for ensuring the quality of the food we eat. At this moment each supermarket refrigerator uses components such as valves, ejectors, compressors... available at each location. This leads to an obvious reduction in transportation costs and is significantly more environmentally friendly, however arises the problematic of designing a different set of controllers for each installation. This thesis tackles this problem by suggesting a reinforcement learning based approach. A variation of the Q-learning algorithm has been designed and implemented, such implementation has been able to optimally control the aforementioned set of systems. The implementation has been made using a recurrent neural network as policy network and such network is able to forecast the rewards with a mean squared error loss of 0.01. Although several challenges have arisen from this approach, it has been proven that indeed it is possible to control unique systems using reinforcement learning provided that enough training data is acquired.