Search

The differential evolution (DE) algorithm is widely regarded as one of the most influential evolutionary algorithms for addressing complex optimization problems. However, the fixed mutation strategy limits the adaptive ability of DE, and the lack of utilization of historical information limits the optimization ability of DE. In this paper, an indicator-based self-learning differential evolution algorithm (ISDE) is proposed. A jump out mechanism based on deep reinforcement learning is adopted to control the mutation intensity of the population. The neural network in the jump out mechanism is designed as a decision maker. The mutation intensity of the population is controlled by the neural network, and the neural network are trained by a double deep Q network algorithm based on the continuous data generated during the evolution process. A population range indicator (PRI) is utilized to describe individual differences in the population. A diversity maintenance mechanism is designed to maintain individual differences according to the value of PRI. The experimental results reveal that the comprehensive performance of ISDE is superior to comparison algorithms on CEC 2017 real-parameter numerical optimization. [ABSTRACT FROM AUTHOR]

The distributed flow shop scheduling problem has become one of the key problems related to the high efficiency impacted factor in the manufacturing industry due to its typical scenarios in real-world industrial applications. In this paper, a knowledge-driven cooperative scatter search (KCSS) is proposed to address the distributed blocking flow shop scheduling problem (DBFSP) to minimize the makespan. The scatter search (SS) is adopted as the basic optimization framework in KCSS. The neighborhood perturbation operator and the Q-learning algorithm are combined to select the appropriate perturbation operator in the search process. Firstly, considering the complexity of distributed scenarios, five search operators are used to construct a disturbance strategy pool. Secondly, the Q-learning algorithm dynamically chooses disturbance strategies to enhance exploration ability and search efficiency. Afterward, a local search method based on neighborhood reconstruction is proposed to perturb the currently found optimal solution to strengthen the ability of KCSS to develop in local areas. In addition, the path relinking mechanism is introduced into the subset combination method to guarantee the diversity of solutions in the optimization process. Finally, the performance of the KCSS algorithm is verified on the benchmark set, and the experimental results demonstrate the robustness and effectiveness of the KCSS algorithm. In addition, 518 of the best-known solutions out of 720 benchmark instances are updated. [ABSTRACT FROM AUTHOR]

Differential evolution (DE) and Estimation of distribution algorithm (EDA) exhibit complementary superiority in solving complex continuous optimization and engineering problems. The design of appropriate strategies coordinated with the two algorithms to balance exploration and exploitation is conducive to obtaining high-precision solutions. A knowledge-driven co-evolutionary algorithm assisted by a cross-regional interactive learning mechanism (KCACIL) is proposed to achieve a comprehensive collaboration between the algorithms, diverse strategies, and cross-regional individuals. Various elite-guided mutation strategies and a self-feedback strategy based on successful experience in light of implicit knowledge are devoted to fulfilling self-learning and cross-regional interactive learning to accomplish individual collaboration and knowledge transfer in the three regions. Reinforcement learning based on ε − g r e e d y and simulated annealing is employed as feedback on the cross-regional individual information to promote the collaboration between opposition-based learning, interaction learning mechanism, and the revised strategy of inferior solutions with small Q values and high distance density. The dynamic self-adaptive adjustment strategies of multiple parameters are adopted to balance diversity and convergence. KCACIL is verified on the CEC 2014, 2017, 2020 benchmark test suites, and engineering applications. Experimental results indicate KCACIL is superior to the state-of-the-art comparison algorithms. • A knowledge-driven comprehensive co-evolutionary algorithm is proposed. • Cross-regional interactive learning with adaptive parameters is introduced. • A collaboration of strategies guided by reinforcement learning is implemented. • The learning mechanisms based on intra-regional implicit knowledge are designed. • Two improvement strategies of solutions specific to the phenomenon are employed. [ABSTRACT FROM AUTHOR]

A co-evolutionary migrating birds optimization algorithm based on online learning policy gradient (CMBO-PG) is proposed to address complex continuous real-parameter optimization problems. In CMBO-PG, a Gaussian estimation of distribution algorithm (GEDA), which enhances the exploitation tendency, is utilized to generate the solutions of the leading flock. The neighborhood solutions of the following flock are produced by a multi-strategy learning mechanism to promote exploration capability. The co-evolution of the leading flock and following flock is realized by the information-sharing mechanism and the operation of destruction and construction to keep the balance of exploration and exploitation. The nonlinear selection of mutation strategies is laborious due to the differences in the ability to address optimization problems. In the mechanism of multi-strategy learning, a long short-term memory (LSTM) is adopted as a selector of mutation strategies to predict the selection probability of three mutation strategies. The evolutionary procedure of the following flock is modeled as a Markov decision process (MDP). The policy gradient (PG) is employed as a model optimizer to control the parameters of LSTM based on the historical feedback information. The performance of CMBO-PG is testified on the CEC 2017 benchmark test suite. The experimental results show that CMBO-PG is superior to the 12 comparison algorithms, including state-of-art algorithms. [ABSTRACT FROM AUTHOR]

The large-scale real value problem in continuous and discrete optimization problems is a challenging issue for researchers and practitioners in the manufacturing domain. An improved artificial bee colony algorithm (ABC) combined with multi-agent reinforcement learning (called MARLABC) is presented for addressing the large-scale real value optimization problem in this study. Two stages including the training and the testing are introduced in the MARLABC via the multi-agent central controller to improve the convergence speed and local exploitation capability of the algorithm. The optimal strategy pool is constructed by training procedures via the multi-agent central controller with Q-learning mechanism. The effective strategy is selected from the optimal strategy pool for each agent during the testing process in the multi-agent central controller. The elite agents in the training population are reserved to generate the testing population to guide the search. The MARLABC algorithm is compared with the ABC variants and state-of-art algorithms on CEC 2017 benchmark problems. The stability and effectiveness of the MARLABC algorithm are confirmed by the experimental results. [ABSTRACT FROM AUTHOR]

A reward function is learned from the expert examples by inverse reinforcement learning (IRL), which is more reliable than an artificial method. The moth–flame optimization algorithm (MFO), which is based on the navigation mechanism of a moth flying at night, has been extensively employed to address the complex optimization problem. An inverse reinforcement learning framework with the Q-learning mechanism (IRLMFO) is designed to strengthen the performance of the MFO algorithm in a large-scale real-parameter optimization problem. The right strategy is chosen by the Q-learning mechanism, using historical data provided by the relevant approach in the strategy pool, which stores strategies that include diverse functions. The competition mechanism is designed to strengthen the exploitation capability of the IRLMFO algorithm. The performance of the IRLMFO is verified on the benchmark test suite in CEC 2017. Experimental results illustrate that the IRLMFO outperforms state-of-the-art algorithms. • An inverse reinforcement learning framework with the Q-learning mechanism is designed. • The framework is applied in the metaheuristic algorithm. • The comparison of IRLMFO with state-of-the-art algorithms is presented. [ABSTRACT FROM AUTHOR]