This work shows different strategies for a Robot to learn the optimal operation of a diverse electrical energy generation system including resources such as thermal, hydroelectric, wind, solar generators and energy accumulators. The large number of variables in these systems results in a huge state space. Thus, computing an explicit representation of the cost function over said space, which is at the heart of most current optimization methods, becomes infeasible. The strategies presented here aim at solving the aforementioned problem by learning an implicit representation of the cost function over the state space. Another key idea is to keep the complexity of the representation at a minimum, in order to obtain a solution which captures the most relevant characteristics of the cost-to-go of the system, with the least possible parameters.