Interactive Preferred Multi-objective Optimization Algorithm Based On Decomposition For High Dimensional Objects

In recent years,the widely application of the multi-objective optimization algorithms in the real life lead to a fast development for this field.Multi-objective optimization algorithms use the advantage of heuristic evolutionary to obtain a series of non-dominate solutions approximating the real Pareto front.However,in the actual application,the user as the final decision-makers,his need may not be the entire Pareto front but only part of the solution or a solution of the Pareto front.By introducing user information into evolution process to guide the algorithm,we can quickly get the user's preferred solutions as well as saving the computing cost.In this paper,we study the multi-objective optimization algorithm for multi-objective optimization of high-dimensional objects and the interactive preference multi-objective optimization algorithm which uses user's information to communicate with the algorithm in the evolution process.Three work points are as follows:1.A multi-objective optimization algorithm based on Maximin operator is proposed.In this algorithm we use the Maximin operator as the selection strategy,which replaces the mechanism of updating the solution using neighborhood information in MOEA/D algorithm.Firstly,the algorithm divides the whole objective space using a set of uniformly distributed weight vectors,and associate one solution for each subspace as the solution under its sub-problem simultaneously.In the selection process,the Maximin operator is used to calculate the individual fitness value,and part of solutions which meet the conditions are chose to use to update of all sub-problems.Due to Maximin operator's characteristic of searching for non-dominated solutions,it can be greatly improve to the convergence performance of the algorithm.Experimental results show that the proposed algorithm has better convergence and diversity on most ZDT and DTLZ problems through a large number of experiments and comparison with other multi-objective optimization algorithms.2.A multi-layer interaction preference multi-objective optimization algorithm through decomposition is proposed.Based on the multi-objective optimization algorithm proposed in the first work,and obtains the user's preference solution through multi-layer interaction is introduced to guide evolution.In the first-layer interaction,we need the user to provide the reference vector and the initial radium,by using the two information,the algorithm can determine a preference range to update the solution within the preference area.Then,we provide the solutions to the user after a predetermined generation,if the user is unsatisfied;the second-layer interaction is needed.The second-layer interaction requires the user to redefine the preference information,and the user is required to select the most satisfactory solution from the results of the first-layer interaction as the new reference vector,also the number of the solutions needed in this interaction is necessary too.After that,our algorithm uses an angle-based strategy to find the solution within the user preference again.In this way,we can narrow the scope of preference gradually,and the user can execute multi-layer interaction and set multiple reference vectors at the same time,meanwhile the user can set all the parameters flexibility therefore our algorithm can obtain the most satisfied solutions for the user.3.A variable range interactive preferred multi-objective optimization algorithm based on decomposition is proposed.First,the user is asked to provide a reference direction and the number of required solutions,then the algorithm searches for the required number of weight vectors closest to the reference direction throughout the objective space as the sub-problems.In the process of interaction,the user enters a parameter of the adjustable preference range and the number of solutions needed in the interactive process.According to the information provided by the user,the algorithm sets two of the update direction for the weight vectors,one is pointing to the reference vector and the other is deviate from it.According to the range parameters and the change of the number of solutions required before and after the interaction,the algorithm can be flexible to change the preference range.