Innovization - Based Evolutionary Large-Scale Multiobjective Optimization

How to handle the high dimensional search space effectively is a burning research question in evolutionary computation when solving large-scale multiobjective optimization problems. To alleviate it, we propose a new evolutionary algorithm with a tactic of innovization-based representation learning, which aims to discover innovative solution principles from compressed representations. Specifically, a neural net, including an encoder and a decoder, is iteratively learned to transform input solutions into the predictions that dominate them. The encoder thus can capture their compressed representations that retain the information necessary for making good predictions. The decoder can create innovative promising solutions from newly explored representations by evolutionary search. Finally, the proposed algorithm is tested on synthetic and real-world benchmarks with the dimension of variables up to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$10^{4}$</tex>. Comparison results show the superior performance of the proposed optimizer in solving these benchmarks within <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$10^{5}$</tex> function evaluations.