Microgrids are a promising tool that can help transition the electricity grid towards a smart grid. They can provide significant benefits to the power grid in the form of increased reliability and flexibility. The microgrid controller sets the electricity consumption and generation of all controllable devices within the microgrid and controls how much electricity is exchanged with the main electricity grid. Model Predictive Control (MPC) has been proposed as a method for developing such microgrid controllers. MPC can deal well with the many different constraints that are imposed on the control of the microgrid. Because controlling a microgrid involves switching devices on and off, microgrids are often modelled as a hybrid system. A downside of MPC for hybrid systems is that, a mixed integer optimization problem must be solved at every control step. Solving mixed integer programs is computationally complex, especially for large scale problems. The difficulty of the optimization problem limits the response time of the microgrid controller. Parametric MPC has been suggested as a way to reduce the computational complexity of the controller and to create an efficient single level MPC controller. In parametric MPC the control input is parametrized according to a set of parameters and a control law. Instead of determining the inputs directly by solving an optimization problem, the optimization problem determines the optimal parameters of the control law. This method can reduce the number of optimization variables significantly and increases the scalability of the controller. To ensure the parametric controller performs well a good parametric control law should be chosen. In this research a new way to determine the parametric control law is proposed. The control law is represented as a combination of expressions. These expressions are represented using a set of expression trees. Using expression trees a wide array of different non-linear functions can be represented. During an offline optimization step, optimal control inputs are determined using a regular MPC controller on a set of scenarios. Expression trees that are able to parametrize the control inputs well are then learned from these control inputs using a genetic algorithm. By using learning methods to determine the control law, the design of the parametric controller can be automated. Using this method there is no need for the control system engineer to determine a parametric control law through trial and error testing. This can speed up the design process and could allow parametric MPC to be used for systems for which input parametrizations are difficult to find. The effectiveness of the proposed approach is illustrated through a case study in which a microgrid is simulated with 2 controllable generators, renewable generation, an uncontrollable load and local energy storage. The performance of the parametric MPC controller determined in the offline optimization is compared with a handcrafted parametric MPC controller and a regular MPC controller. The estimated economic cost of operating the simulated microgrid using the different controllers is compared. Furthermore the computational complexity of the different controllers is analysed. The results show that the offline trained parametric controller achieves similar performance in both operating cost and computational complexity as the handcrafted parametric controller. This shows that the proposed offline optimization algorithm can be used to determine an effective control law for a parametric MPC controller. This will make it easier to design parametric MPC controllers for different systems in the future.