Learning-based multi-objective optimization through ANN-assisted online Innovization
Sukrit Mittal, Dhish Kumar Saxena, Kalyanmoy Deb · 2020
Learning for effective problem information from early iterations in an optimization run and utilizing it to improve convergence properties has presented important directions in evolutionary computation (EC) research. In this paper, we propose an ANN-assisted modeling approach which learns from pairs of solutions - the improvements in the variable space, and uses the resulting ANN as an innovized repair operator in an iterative manner. Although the concept can be easily applied to single-objective problems, in this paper, we develop the overall procedure for multi-objective optimization. On a number of test and engineering problems involving two and three-objective problems, we demonstrate a faster execution in terms of the hypervolume metric. The results are encouraging and motivate further exploration to more challenging problems.