Many-objective Optimization Based On Evolutionary Multi-tasking And Transfer Learning

In the real world,many problems can be abstracted as optimization problems.Evolutionary algorithm(EA)is widely used to solve multi-objective problems(MOPs)due to its population-based feature which allows individuals to simultaneously approximate different parts of the Pareto front within a single execution.For many-objective optimization problems(MaOPs),there are more than three objectives to be optimized.O ne major reason behind the failure of most MO EAs in solving MaOPs is that the selection criterion based on the Pareto dominance fails to distinguish solutions since most of the solutions are non-dominated.Besides,constraints are always introduced into practical problems.We propose two algorithm to solve MaOPs with and without constrains.The main work of this paper outlines as follows:Firstly,due to the ineffectiveness of Pareto dominance in many-objective optimization,we introduce multi-agent mechanism.Each agent performs the task of guiding the subpopulation to evolve towards the true Pareto front in the direction of the specified reference vector.Owing to the autonomous behavior of the agent,the scope of selecting individuals is limited within the agent,and the fitness of individuals is a scalar criteria which combines convergence and diversity.The environmental selection method could distinguish individuals effectively without relying on Pareto dominance.Secondly,in order to alleviate the imbalance of convergence and diversity,we use the diversity archive to interact with multi-agent mechanism.Since the reference vectors are evenly distributed in the objective space,the population can maintain convergence and diversity simultaneously during the process that agents guide the subpopulations to evolve towards the true Pareto front.Diversity archive can not only improve the diversity of the population,but also significantly strengthen the convergence by executing the transfer of good individuals between different tasks through aggregation and redistribution among agents.Finally,when solving constrained multi-objective optimization problems,an important issue is how to balance convergence,diversity and feasibility simultaneously.To address this issue,we propose an ensemble constrained-handling techniques based many-objective evolutionary algorithm.In this algorithm,we design a self-adaptive penalty mechanism to penalize those solutions that violate the constrained according to the feasibility proportion of the current population.The algorithm which synthesizes the self-adaptive penalty mechanism,constrained dominance principle,multi-agent mechanism and diversity archive could guide the population to evolve through the infeasible domain and approach to the true Pareto front.