Learning Descriptors for Novelty-Search Based Instance Generation via Meta-evolution

Figure 1: Meta-Learning of Descriptors: (1) CMA-ES is used to evolve the weights of an NN with fixed architecture that projects a feature-vector to 2D. (2) For each evolved NN (weight vector), a Novelty Search (NS) algorithm is run to evolve knapsack instances that are discriminatory w.r.t. a solver portfolio and diverse in the 2D projection defined by the corresponding NN.(3) The output of the NS algorithm is a set of instances which are projected into a fixed 8D hypercube.The coverage of the hypercube is assigned as the fitness of the corresponding NN weight vector.(4) CMA-ES repeats the above procedure for a fixed number of generations.