Search

Differential evolution is one of the most powerful stochastic real-parameter optimization algorithms currently, and its performance depends heavily on control parameters and mutation strategy. In recent years, methods to select favorable parameters control and mutation strategy when solving various optimization problems have attracted increasing attention. To choose an appropriate mutation strategy and control parameters for a given optimization problem, in this paper, a Adaptive Differential Evolution Algorithm Based on Deeply-Informed Mutation Strategy and Restart Mechanism (ADEDMR) is proposed, and the ADEDMR algorithm has the following characteristics: First, a deeply-informed mutation strategy is proposed, which takes into account the information of suboptimal solutions discarded by selection and inherits the advantages of the powerful "DE/target-pbest/1/bin", aiming to obtain a better perceptual landscape of the target function and improve the candidate diversity of the trial vector. Second, according to the evolution process, the segmentation method is used to control F , which alleviates the scaling of F in the wrong direction, and makes the newly generated F fit more accurately. Third, a new population restart mechanism is adopted to further enhance population diversity by adaptively enhancing the search ability of hopeless individuals and randomly replacing some inferior individuals with wavelet walks. To evaluate the performance of our proposed algorithm, comparative experiments are conducted on 72 benchmark functions from the CEC2014, CEC2017 and CEC2022 test suites. Experimental results show that the proposed ADEDMR has higher convergence accuracy, better optimization ability when solving high-dimensional complex functions, and is competitive with six recent strong DE variants. [ABSTRACT FROM AUTHOR]

In this paper, we proposed a novel fitness-value-independent Differential Evolution to tackle complex real-parameter single-objective optimization. There were three innovations in the algorithm: First, dimension improvements based adaptation schemes for control parameters F and C R were advanced in this paper. Different from those fitness value improvements based parameter control in the recent winner DE variants, our algorithm has fitness-value-independent characteristic, therefore, it can be applied in much wider optimization scenarios especially for those that the exact fitness values of the objectives are unavailable. Second, a combined parabolic–linear reduction scheme for population size is employed in the algorithm based on the observation that the slower reduction of population size at the earlier stage of the evolution usually helps to get better perception of the objectives while a linear reduction of population size in the later stage helps to get better exploitation. Third, an indicator is proposed to monitor the diversity of the population and a corresponding population enhancement technique is launched when the diversity is detected bad. The first two innovations have the P arameter a daptive characteristic of the DE algorithm while the third innovation refers to p opulation e nhancement t echnique, therefore, we name our algorithm "the PaDE-pet algorithm". Then, the algorithm is validated under a larger test suite containing 88 benchmarks from CEC2013, CEC2014 and CEC2017 test suites for real-parameter single objective optimization, which may avoid over-fitting problem in comparison with employing a test suite containing a small number of benchmarks. The experiments support the superiority of the novel PaDE-pet algorithm in comparison with several recently proposed powerful DE variants. • A black-box model reflecting complex optimization without available fitness values. • The model reveals the fitness-value-dependent weakness of recent winner DE variants. • A novel fitness-value-independent DE algorithm was proposed to tackle the model. • New adaptation schemes for control parameters and population enhancement technique. • To avoid over-fitting, a larger test suite is employed in algorithm validation. [ABSTRACT FROM AUTHOR]