Learning to Learn Evolutionary Algorithm: A Learnable Differential Evolution

Abstract

Research on evolutionary optimization has flourished for several decades. Now it has come to a turning point. With the advancement of artificial intelligence, especially deep learning and reinforcement learning, it is becoming appealing to rethink the design and development of evolutionary algorithm (EA). From our perspective, a new-generation EA should be <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">learned</i> rather than manually designed, based on learning from optimization experiences (such as obtained from optimizing a family of optimization problems), the deep understanding of the roles of recombination operators, and the usage of experiences extracted through history optimization trajectories, so as to intelligently decide the control parameters that can adapt to the problem landscape changes. This learning can be conducted by strongly coupling with reinforcement learning since an evolutionary search procedure can be modeled as a Markov Decision Process (MDP). In this paper we propose a framework for automatic learning of EA, and present an exemplar study on learning a differential evolution (DE). Experimental results show that the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">learned</i> adaptive DE is very competitive to some recent EAs on a commonly-used test suite, which indicates that the proposed learning framework has a great potential for the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">automatic</i> design of promising EAs.