Discovering a Single Neural Network Controller for Multiple Tasks with Evolutionary Algorithms

(1) Background: Multi-Objective Optimization is a prominent research area, in which approaches for the simultaneous solution of multiple objectives are proposed. The possibility to discover a set of parameters optimizing all the goals can be achieved only if the considered problems are rather trivial, while compromise solutions are generally discovered. Things become even more complex when the set of parameters is used in opposite, and potentially conflicting, ways. (2) Methods: in this work, we compared some state-of-the-art Evolutionary Algorithms with regard to the optimization of different conflicting objectives, by highlighting strengths and weaknesses of the different approaches. In particular, we considered four benchmark problems — 4-bit parity, double-pole balancing, grid navigation and test function optimization — to be solved simultaneously. (3) Results: our investigation identifies the algorithms leading to a better optimization. Moreover, we provide an analysis of the solutions to the single objectives, hence illustrating how the different methods address the problems. (4) Conclusions: two algorithms emerge as the most suitable methods for dealing with the considered scenario. Notably, a relatively simple strategy is not significantly inferior to a more sophisticated one, thus emphasizing that there exists a non-negligible relationship between problems and algorithms, and discovering general methods is extremely difficult.