Multi-Objective Optimization Algorithm With Preference Via Active Learning

In the field of engineering and economics,we often encounter optimization problems that need to consider multiple objectives at the same time.This kind of problem is called multi-objective optimization problem(MOP).The objective functions of MOP often conflict with each other,and the improvement of a certain objective often leads to the decline of one or more other objectives.The solution to MOP is a set of solutions that compromise on each goal.Evolutionary multi-objective optimization(EMO)is a successful application of evolutionary computation(EC)in the field of multi-objective optimization problems(MOPs).Its mechanism based on population search and the principle of survival of the fittest can obtain a set of optimal solutions of MOP in a single run.EMO can still obtain global approximated optimal solutions of MOP without making additional assumptions about the problem.Above advantages make EMO a powerful way to solve MOPs.In order to enable traditional EMO to obtain the optimal solutions preferred by the decision maker(DM),interactive evolutionary multi-objective optimization(IEMO)is proposed.IEMO can combine preference learning and optimize while learning in an interactive manner,so as to guide the population to focus on searching the region of interest(ROI)of the DM in the objective space.However,in the interaction process,the traditional IEMO method only uses random sampling to select a set of candidate solutions to be judged by the DM according to her or his preference,without considering the preference model and the existing preference information itself.In order to select the most informative candidate solutions that can improve the performance of the preference model,this thesis adopts the thought of active learning to model the preference of the DM as a preference ranking model and dynamically and actively selects candidate solutions for the DM to express her or his preference information in the interaction process,so as to obtain the most informative preference information.In this way,the cognitive load of the DM in the interaction process can be significantly reduced.Based on the preference ranking model,this paper proposes an interactive evolutionary multi-objective optimization framework IEMO/AR(Interactive Evolutionary Multiobjective Optimization via Active Ranking),which can be embedded in by any type of evolutionary multi-objective optimization algorithms(EMOAs).Using this framework,interactive evolutionary multi-objective optimization algorithms for low and high dimensional multi-objective optimization problems are designed respectively.The performance of the algorithm is compared with several classical algorithms in many aspects on a large number of test problems.Through the summary and analysis of the experimental results,it is verified that the proposed algorithm can obtain the high-quality optimal solutions in line with the preference of the DM when facing MOP.