Adaptive Operators for Genetic Programming to Identify Optimal Energy Flow Controllers

Genetic programming can find nearly optimal solutions for complex problems like minimizing a building's energy costs by optimally controlling its energy flows. For such problems, usually multiple controllers are necessary. In order to allow a faster convergence in combination with a more fine-grained and directed search, this work presents new adaptive crossover and mutation operators. Instead of applying the operators always to all symbolic regression trees in a solution candidate, the new operators are applied to all trees only in the beginning and then to a randomly chosen group of them as soon as a threshold is reached. Towards the end of the training, the adaptive operators then switch to applying crossover and mutation to only one of the trees in a solution candidate for a more fine-grained search. Additionally, a new crossover is proposed where the children solution candidates are themselves evaluated for their performance before promoting one of them to the next generation in order to assure a more directed search. To evaluate these new operators, a total of twelve energy management controllers is trained with the Offspring Selection Genetic Algorithm and are evaluated for training results in form of the needed number of evaluated solutions and generations as well as their ability to reduce the energy costs and their learned behaviour. Results show that the proposed adaptive operators achieve very similar results to the baseline optimization and that the Best Child crossover is the fastest to converge.