Multi tree operators for genetic programming to identify optimal energy flow controllers

Genetic programming is known to be able to find nearly optimal solutions for quite complex problems. So far, the focus was more on solution candidates that hold just one symbolic regression tree. For complex problems like controlling the energy flows of a building in order to minimize its energy costs, this is often not sufficient. This is why this work presents a solution candidate implementation in HeuristicLab where they hold multiple symbolic regression trees. Additionally, also new crossover and mutation operators were implemented as the existing ones cannot handle multiple trees in one solution candidate. The first type of operators applies them on all trees in the solution candidate, whereas the second one only applies them to one of the trees. It is found that applying the mutator to only one of the trees significantly reduces the training duration. Applying the crossover to one of the trees instead of all needs longer training times but can also achieve better results.