Surrogate-assisted evolutionary algorithms for expensive optimization problems have gained considerable attention in recent years. In many real-world optimization problems, we may face expensive optimization problems with multiple optimal solutions. Locating multiple optima for such expensive problems is qualitatively challenging. This study proposes a surrogate-assisted differential evolution based on region decomposition to seek multiple optima for expensive multimodal optimization problems. In this study, we have designed three major components: 1) the adaptive region decomposition, 2) the multilayer perceptron-based global surrogate, and 3) the self-adaptive gradient descent-based local search. First, the improved adaptive region decomposition detects promising subregions at the beginning phase, and continuously discards inferior subregions successively. Second, the multilayer perceptron-based surrogate and self-adaptive gradient-based mutation work in a collaborative manner on distinct sub-populations to seek multiple optimal solutions. Overall, an attempt has been made to solve expensive multimodal optimization problems. Systematic experiments on 20 test functions show the encouraging and promising performance of our proposed approach.
Solving large-scale, constrained, and non-linear optimization problems is crucial for the Internet of Things (IoT) due to its wide range of real-life applications. However, there is no unified approach for handling constraints and optimizing objective functions. This paper proposes a tri-objective general framework (TriGF) and an efficient differential evolution method enhanced with adaptive gradient-based mutation (AGM), termed AGM-DE. Within the TriGF, AGM-DE explores the entire feasible region by considering both constraints and the objective function. The goal is to achieve global optimality and fast convergence for the self-assembly of satellite IoT devices under constraints. AGM is an adaptive refinement technique that uses gradient information to reduce the search space and speed up optimization. In our AGM approach, we incorporate gradient information from the objective function to mitigate the negative effects of classic constraint-based gradient descent and reduce its inherent greediness. To validate AGM-DE’s effectiveness, we conducted extensive simulations on 57 benchmark problems with diverse dimensions and constraints. The results demonstrate AGM-DE’s exceptional ability to manage constraints in 56 of these 57 test functions, outperforming five leading methods in optimization efficacy and consistency. We also assessed AGM-DE’s application in optimizing IoT device self-assembly within a satellite layout, subject to heat pipe constraints. Comparative analyses highlight AGM-DE’s robustness and superior search capabilities in deriving layout schemes. Remarkably, these schemes outperform existing best-known solutions for IoT configurations involving 40 to 90 nodes with 80 to 180 variables, confirming AGM-DE’s suitability for a wide range of large-scale constrained IoT challenges.
The presence of equality and inequality constraints is of significant importance in expensive constrained optimization. This study aims to develop a surrogate-assisted evolutionary algorithm capable of addressing three types of constraints commonly encountered in expensive optimization problems: 1) those with only inequality constraints, 2) those with only equality constraints, and 3) those with both equality and inequality constraints. To achieve this, a radial basis function neural network is employed as a surrogate model to approximate expensive objective functions and individual constraints. Leveraging a consensus measure among all individual constraints, an improved infill sampling criterion is proposed to identify the most promising candidates. In addition, a hybrid local search strategy refines infeasible yet promising solutions via both surrogate-driven evolution and a model-free gradient-based mutation. To balance these search modes cost-effectively, a stagnation strategy enables adaptive switching between exploration and exploitation. Experimental results on 58 test instances demonstrate the efficacy of the proposed approach across all three constraint types under a strict budget of 1000 expensive function evaluations. These findings validate the methodology and show competitive performance relative to state-of-the-art methods.