ADAPTIVE SELECTION OF AUXILIARY OBJECTIVES IN MULTIOBJECTIVE EVOLUTIONARY ALGORITHMS

• Main Authors: I. A. Petrova, A. S. Buzdalova, A. A. Shalyto
• Format: Article
• Language: English
• Published: Saint Petersburg National Research University of Information Technologies, Mechanics and Optics (ITMO University) 2016-05-01
• Series: Naučno-tehničeskij Vestnik Informacionnyh Tehnologij, Mehaniki i Optiki
• Subjects: evolutionary algorithm; reinforcement learning; multiobjective optimization; travelling salesman problem
• Online Access: http://ntv.ifmo.ru/file/article/15505.pdf

Subject of Research.We propose to modify the EA+RL method, which increases efficiency of evolutionary algorithms by means of auxiliary objectives. The proposed modification is compared to the existing objective selection methods on the example of travelling salesman problem. Method. In the EA+RL method a reinforcement learning algorithm is used to select an objective – the target objective or one of the auxiliary objectives – at each iteration of the single-objective evolutionary algorithm.The proposed modification of the EA+RL method adopts this approach for the usage with a multiobjective evolutionary algorithm. As opposed to theEA+RL method, in this modification one of the auxiliary objectives is selected by reinforcement learning and optimized together with the target objective at each step of the multiobjective evolutionary algorithm. Main Results.The proposed modification of the EA+RL method was compared to the existing objective selection methods on the example of travelling salesman problem. In the EA+RL method and its proposed modification reinforcement learning algorithms for stationary and non-stationary environment were used. The proposed modification of the EA+RL method applied with reinforcement learning for non-stationary environment outperformed the considered objective selection algorithms on the most problem instances. Practical Significance. The proposed approach increases efficiency of evolutionary algorithms, which may be used for solving discrete NP-hard optimization problems. They are, in particular, combinatorial path search problems and scheduling problems.