Research And Application On Constrained Multi-objective Evolutionary Algorithm Based On Reinforcement Learning

In most situations,continuous numerical optimization problems and real-world engineering optimization problems are constrained multi-objective optimization problems(CMOP).Evolutionary algorithm is a kind of heuristic intelligent algorithm based on group search and individual information exchange set.When solving constrained optimization problems,evolutionary algorithms have the characteristics of parallelism and robustness.It is very suitable for solving complex and nonlinear problems.The performance of evolutionary algorithms deeply depends on their strategies of mutation and crossover.These strategies consist of some control parameters,which can affect the performance of the algorithms.When solving CMOP,the parameters of constraint handling mechanism must be considered.However,the parameters are always preset in many algorithms.Compared with algorithms that have preset control parameters,the algorithms with adaptive method to control parameters are more superior to adapt different problem scenarios.This paper presents a constrained multi-objective evolutionary algorithm called RLMOEA/D-ε,which is designed based on reinforcement learning(RL).The algorithm uses multi-objective evolutionary algorithm based on decomposition(MOEA/D)with the constraint handling mechanism(MOEA/D-ε)as the basic framework.It uses long short term memory(LSTM)as the trained network.It designs a reward feedback mechanism based on inverted generational distance(IGD)value,which can adaptively adjust the parameters of the differential evolution operators and the constraint relaxation factor.The algorithm uses the differential evolution operator to obtain new individuals,and applies the constraint relaxation factor to naturally keep evolving until termination condition is met.The suggested algorithm effectively solves the performance problems which are caused by their sensitive parameters.It can effectively keep the balance of convergence,diversity and feasibility.Experimental results show that the convergence and diversity of RL-MOEA/D-ε significantly better than the other four classical constrained multi-objective evolutionary algorithms in LIR-CMOP3-14 test problem sets.In solving LIR-CMOP1 and LIR-CMOP2,although the performance of RL-MOEA/D-εε is slightly worse than MOEA/D-εSR,it is better than other constrained multiobjective evolutionary algorithms.The simulation results also show that the performance RLMOEA/D-εε is significantly superior to other constrained multi-objective evolutionary algorithms in solving I-beam optimization problem,which presents RL-MOEA/D-ε is competitive in dealing with practical engineering problems.