Memetic Computing最新文献

Abstract

Deep neural networks, specifically deep convolutional neural networks (DCNNs), have been highly successful in machine learning and computer vision, but a significant challenge when using these networks is choosing the right hyperparameters. As the number of layers in the network increases, the search space also becomes larger. To overcome this issue, researchers in deep learning have suggested using deep compression techniques to decrease memory usage and computational complexity. In this paper, we present a new approach for compressing deep CNNs by combining filter and channel pruning methods based on Evolutionary Algorithms (EA). This method involves eliminating filters and channels in order to decrease the number of parameters and computational complexity of the model. Additionally, we propose a bi-level optimization problem that interacts between the hyperparameters of the convolution layer. Bi-level optimization problems are known to be difficult as they involve two levels of optimization tasks, where only the optimal solutions to the lower-level problem are considered as feasible candidates for the upper-level problem. In this work, the upper-level problem is represented by a set of filters to be pruned in order to minimize the number of selected filters, while the lower-level problem is represented by a set of channels to be pruned in order to minimize the number of selected channels per filter. Our research has focused on developing a new method for solving bi-level problems, which we have named Bi-CNN-Pruning. To achieve this, we have adopted the Co-Evolutionary Migration-Based Algorithm (CEMBA) as our search engine. The Bi-CNN-Pruning method is then evaluated using image classification benchmarks on well-known datasets such as CIFAR-10 and CIFAR-100. The results of our evaluation demonstrate that our bi-level proposal outperforms state-of-the-art architectures, and we provide a detailed analysis of the results using commonly employed performance metrics.

Graphical abstract

Portfolio optimization will apply the concept of diversification across asset classes, which means investing in a wide variety of asset types and classes for a risk-mitigation strategy. Portfolio optimization is a way to maximize net gains in a portfolio while minimizing risk. A portfolio means investing in a wide variety of asset types and classes for a risk-mitigation strategy by the investor. In this paper, factor analysis and cluster algorithm are used to screen stocks and an improved differential evolution algorithm for solving portfolio optimization model is proposed. By comprehensively analyzing the stock data with factor analysis and k-means clustering algorithm, it has found that important factors have important effect on stock price movement, and finally 10 stocks are selected with investment value. Besides, a Mean-Conditional Value at Risk (CVaR) model is constructed, which takes into account both the cost function and the diversification constraint. Finally, a second-order memetic differential evolution (SOMDE) algorithm is presented for solving the proposed model. The experiments show that the proposed SOMDE algorithm is valid for solving the Mean-CVaR model and that factor analysis for stock selection can benefit portfolio with higher return and less risk greatly.

Recently, surrogate-assisted multi-objective evolutionary algorithms have achieved much attention for solving computationally expensive multi-/many-objective optimization problems. An effective infill sampling strategy is critical in surrogate-assisted multi-objective evolutionary optimization to assist evolutionary algorithms in identifying the optimal non-dominated solutions. This paper proposes a Kriging-assisted many-objective optimization algorithm guided by two infill sampling criteria to self-adaptively select two new solutions for expensive objective function evaluations to improve history models. The first uncertainty-based criterion selects the solution for expensive function evaluations with the maximum approximation uncertainty to improve the chance of discovering the optimal region. The approximation uncertainty of a solution is the weighted sum of approximation uncertainties on all objectives. The other indicator-based criterion selects the solution with the best indicator value to accelerate exploiting the non-dominated optimal solutions. The indicator of an individual is defined by the convergence-based and crowding-based distances in the objective space. Finally, two multi-objective test suites, DTLZ and MaF, and three real-world applications are applied to test the performance of the proposed method and four compared classical surrogate-assisted multi-objective evolutionary algorithms. The results show that the proposed algorithm is more competitive on most optimization problems.

Βrain storm optimization (BSO) is a swarm-intelligence clustering-based algorithm inspired by the human brainstorming process. Electromagnetism-like mechanism for global optimization (EMO) is a physics-inspired optimization algorithm. In this study we propose a novel hybrid metaheuristic evolutionary algorithm that combines aspects from both BSO and EMO. The proposed algorithm, named EMotion-aware brain storm optimization, is inspired by the attraction–repulsion mechanism of electromagnetism, and it is applied in a new emotion-aware brainstorming context, where positive and negative thoughts produce ideas interacting with each other. Novel contributions include a bi-polar clustering approach, a probabilistic selection operator, and a hybrid evolution process, which improves the ability of the algorithm to avoid local optima and convergence speed. A systematic comparative performance evaluation that includes sensitivity analysis, convergence velocity and dynamic fitness landscape analyses, and scalability assessment was performed using several reference benchmark functions from standard benchmark suites. The results validate the performance advantages of the proposed algorithm over relevant state-of-the-art algorithms.

This paper proposes top-level dual-exploitation particle swarm optimization (TLDEPSO), which aims to use the evolutionary experience between particles better and enhance the convergence performance of the algorithm. In TLDEPSO, the population is divided into top-level particles and ordinary particles according to fitness, and each iteration is divided into two stages to be executed. For the first stage, a particle modification method based on gene editing technology is proposed and applied to top-level particles to improve the search direction of the population and explore the problem space better. For other ordinary particles in the population, the learning strategy of the canonical ring neighborhood topology PSO is used to update the velocity and the position to maintain the diversity of the population. For the second stage, a top-level neighborhood exploration mechanism is proposed for top-level particles to accelerate the algorithm’s convergence. In addition, an adaptive dynamic adjustment mechanism for the parameters of acceleration coefficient, inertia coefficient and the number of top-level particles is proposed to balance better the global exploration and local exploitation capabilities of the algorithm. On the latest CEC2022 test benchmark, comparison and statistical analysis with seven advanced PSO variants and three CEC competition top algorithms demonstrate TLDEPSO’s superior performance in solving functional problems with different fitness landscapes.

Parameter pruning is one of the primary methods for compressing CNN models, aiming to reduce redundant parameters, the complexity of time and space, and the calculation resources of the network, all while ensuring minimal loss in the network’s performance. Currently, most existing parameter pruning methods adopt equal pruning rates across all layers. Different from previous methods, this paper focuses on the optimal combination of each layer’s pruning rates within a given pruning rate of the whole model. Genetic algorithm is used to determine the pruning rate for each layer. It’s worth noting that while the pruning rate for individual layers may vary, the average pruning rate across all layers does not exceed the given pruning rate. Experimental validation is conducted on CIFAR10 and ImageNet ILSVRC2012 datasets using VGGNet and ResNet architectures. The results show that the accuracy loss and the FLOPs of the pruned model using our method are superior to those pruned using previous methods.