Plastic Fitness Predictors Coevolved with Cartesian Programs

fitness predictors, cartesian genetic programming, coevolution, phenotypic plasticity

Coevolution of fitness predictors, which are a small sample of all training data for a particular task, was successfully used to reduce the computational cost of the design performed by cartesian genetic programming. However, it is necessary to specify the most advantageous number of fitness cases in predictors, which differs from task to task. This paper proposes to introduce a new type of directly encoded fitness predictors inspired by the principles of phenotypic plasticity. The size of the coevolved fitness predictor is adapted in response to the phase of learning that the program evolution goes through. It is shown in 5 symbolic regression tasks that the proposed algorithm is able to adapt the number of fitness cases in predictors in response to the solved task and the program evolution flow.