For optimization in the smart grid, distributed algorithms based on decoders for handling individual constraints of different energy resources are a promising approach to tackle the scalability and versatility of controlled devices. Decoders based on machine learning can capture the operational capabilities and serve as a means for systematically ensuring the feasibility of solution candidates during optimization. Currently, decoders are trained based on a training set for a specific initial state predicted for the start time of the optimization period. Thus, a new decoder has to be trained for any new initial operational state of the energy resource. This paper explores a new approach based on Cartesian genetic programming to train a decoder that can be parameterized with different initial states. We train decoders for co-generation plants with a range of different states of charge for the thermal buffer store and demonstrate that such decoders can be obtained in a reasonable training time and with sufficiently good performance over the whole range of temperatures.