Xinchao Zhao

Clustering and pattern search for enhancing particle swarm optimization with Euclidean spatial neighborhood search

There are many well-known particle swarm optimization (PSO) algorithms which consider ring/star/von Neumann et al. topological neighborhood and scarcely aim at Euclidean spatial neighborhood structure. k-Nearest Neighbors (k-NN) is a kind of clustering method to find the necessary representatives among a group of objects efficiently. Pattern search (PS) is a successful derivative-free coordinate search method for global optimization. All these observations inspire the innovative ideas to propose an enhanced particle swarm optimization algorithm (pkPSO). Particles efficiently explore for the promising areas and solutions with clustering on the Euclidean spatial neighborhood structure. Particle swarm continuously exploits at the just found promising areas with PS strategy at the latter stage of optimization. The cooperative effect of k-NN and PS strategies is firstly verified. Based on classical, rotated and shifted benchmarks, extensive experimental comparisons indicate that pkPSO has a competitive performance when comparing with the well-known PSO variants and other evolutionary algorithms.

Discover Community Leader in Social Network with PageRank

Community leaders are individuals who have huge influence on social network communities. Discovering community leaders in social networks is of great significance for research on the structures of the social networks and for commercial application. Based on the core idea of the PageRank algorithm, this paper firstly processes data selected from Sina microblog, and extracts three key indicators, comprising the number of followers, the number of comments and the number of reposts; then based on their mutual relationship, that is following or followed, it obtains the weight of influence for each individual user; and then after a finite number of iterations, this paper identifies the community leader in Sina microblog, by which its comprehensive influence on its community are reflected.

New modified bare-bones particle swarm optimization

Bare-bones Particle Swarm Optimization (BPSO) is a simplified PSO variant, which has shown potential performance on many multimodal optimization problems. However, BPSO is also possible to be trapped into local optima for high-dimensional and complicated optimization problems. In order to enhance the performance of BPSO, this paper presents a modified BPSO, called NMBPSO. It combined the ideas of the traditional PSO and a modified BPSO to improve the capacity of balancing exploration and exploitation during the search process. To verify the effect and benefit of the proposed algorithm, a set of well known benchmark functions are employed and compared against some competitive PSO variants. Experiment results indicate that NMBPSO performs better than the traditional PSO, BPSO and a modified BPSO algorithm.

Cooperative Co-evolution with Principal Component Analysis for Large Scale Optimization

This paper attempts to address the problem of large scale optimization and high dimensional optimization using principal component analysis (PCA) strategy with differential evolution (DE) based on Cooperative Co-evolution (CC) framework. Decomposition problem is a major obstacle for large-scale optimization problems. The aim of this paper is to propose effective dimension decomposition method of PCA strategy for capturing the main information among dimensions. PCA strategy can measures most of the contribution information of dimension and uses it for identifying main dimension to guide them to group the most promising subcomponents in CC framework. Then each subcomponents can be solved using an evolutionary optimizer to find the optimum values. The experimental results show that this new technique is more effective than some existing grouping methods.

Origin Illusion, Elitist Selection and Contraction Guidance

Most of existing swarm intelligence (SI) algorithms is modeling based on natural phenomena. Firstly, different from the previous practices, this paper constructs a mathematical model based on the traditional optimization algorithms. To simplify this model, a new algorithm Linear Transformation and Elitist Selection algorithm (LTES) is proposed. Experiment shows that the algorithm has origin illusion phenomenon. Then, this paper observes origin illusion phenomenon for the population-based optimization algorithm, and experiments shows that crossover operator is an effective way for LTES’ origin illusion problem. Finally, another algorithm Contraction and Guidance Algorithm (CGA) is proposed to prove that elitist selection is not necessary. The experimental results show that both algorithms are effective.

An Orthogonal Genetic Algorithm with Multi-parent Multi-point Crossover for Knapsack Problem

According to the inherent feature of knapsack problem, a multi-parent multi-point crossover operation (MP2X) is proposed, which is implanted with orthogonal experimental design method. The aim of implementing orthogonal experimental design method to MP2X operation is to fully utilizing the inherent information from multiple component of multiple individuals. Based on MP2X operation and orthogonal design method, a genetic algorithm variant (MPXOGA) is proposed in this paper. The simulation results on classic knapsack instances show that MPXOGA is better than several other solvers, including Hybrid Genetic Algorithm (HGA), Greedy Genetic Algorithm (GGA), Greedy Binary Particle Swarm Optimization Algorithm (GBPSOA) and Very Greedy PSO (VGPSO) in the ability of finding optimal solution, the efficiency and the robustness.

A new multi-stage perturbed differential evolution with multi-parameter adaption and directional difference

A new multi-stage perturbed differential evolution (MPDE) is proposed in this paper. A new mutation strategy “multi-stage perturbation” is implemented with directivity difference information strategy and multiple parameters adaption. The DE/current-to-pbest is introduced to increase the population diversity while remaining its elitist learning behavior in this architecture. The multi-stage perturbation-based mutation operation utilizes the Normal random distribution with adjustable variance to perturb the chosen solutions. Multiple parameters are adaptively adjusted to appropriate values to match the current search status of algorithm. It is thus helpful to enhance the performance and the robustness of algorithm. Simulation results show that the newly proposed MPDE is better than, or at least comparable to CLPSO, SPSO2011, NGHS, jDE, CoDE, SaDE and JADE algorithms in terms of optimization performance based on CEC2015 benchmark function.

Conservatism and Adventurism in Particle Swarm Optimization Algorithm.

Particle Swarm Optimization (PSO) is a widely used optimization algorithm in industrial and academic fields. In this paper, three improved PSO variants are proposed. The main ideas of them are that a coefficient v is added to control the velocity augment of particles to the new position on different dimension. The first one is under the guidance of conservatism which is an inspiration of Differential Evolution (DE), namely, particles preserve more information from their previous positions and move in a smaller search space. This algorithm shows that particles are possible to escape from the current neighborhood and for promising search area if they take more previous information. The second one is guided by adventurism for better exploration, which means a larger search space to particles. The third one can be considered as a compromise between conservatism and adventurism. This algorithm shows that a balanced cooperation with a little conservative in more adventures will make PSO more competitive. Experimental results show that the proposed strategies of all the three algorithms are effective based on CEC2015 benchmarks. All of them are better than the traditional PSOs and the third improved variant performs better than all the other competitors.

Elite-Leading Fireworks Algorithm

Fireworks algorithm (FWA) is effective to solve optimization problems as a swarm intelligence algorithm. In this paper, the elite-leading fireworks algorithm (ELFWA) is proposed based on dynamic search in fireworks algorithm (dynFWA), which is an important improvement of FWA. In dynFWA firework is separated to two group: core-firework (CF) and non-core fireworks (non-CFs). This paper takes some beneficial information from non-CFs to reinforce the local search effect of CF. Random reinitialization and elite-leading operator are used to maintain the diversity of the non-CFs, which play an important role in global search. Based on the CEC2015 benchmark functions suite, ELFWA has a very competitive performance when comparing with state-of-the-art fireworks algorithms, such as dynFWA, dynFWACM and eddynFWA.

A Recommendation Approach Based on Latent Factors Prediction of Recurrent Neural Network

Recommender systems have received much attention due to their wide applications. Current recommender approaches typically recommend items to user based on the rating prediction. However, the predicted ratings cannot truly reflect users interests on items because the rating prediction is usually based on history data and does not consider the effect of time factor on uses interests (behaviors). In this paper, we propose a recommendation approach combining the matrix factorization and a recurrent neural network. In this approach, all the items rated by a user are considered as time series data. The matrix factorization is used to obtain latent vectors of those items. The recurrent neural network is taken as a time series prediction model and trained by the latent vectors of historical items, and then the trained model is used to predict the latent vector of the item to be recommended. Finally, a recommendation list is formed by mapping the latent vector into a set of items. Experimental results show that the proposed approach is able to produce an effective recommend list and outperforms those comparative approaches.