Neural Computing and Applications

The flower pollination algorithm (FPA) is a relatively new natural bio-inspired optimization algorithm that mimics the real-life processes of the flower pollination. Indeed, this algorithm is based globally on two main rules: the global pollination (biotic and cross-pollination) and the local pollination (abiotic and self-pollination). The random permutation between these latter allows to keep a permanent balance between intensification and diversification. However, this procedure causes an involuntary orientation toward a bad solution (local optima). In addition, FPA illustrates an inadequacy in terms of intensification and diversification of new solutions; this has become clear when the complexity of the treated problem is increased. Further, FPA has also another insufficiency, which is its slow convergence rate caused in principle by its weak intensification. In this paper, to overcome these weaknesses, we have introduced some modifications on the basic FPA algorithmic structure based on the two following improvements: (1) Generating a set of global orientations (toward global or local pollination) for all members of the population. Indeed, each element (global orientation) in this set is composed of a fixed number (equal to the population size) of sub-random orientation. Thus, the number of elements is fixed by the designer, which enhances significantly the diversification characteristic. (2) Constructing a set of best solution vectors relating to all generated global orientations. In fact, this set is compared at each iteration to a fixed number of actual solution vectors to select the best among them based on their fitness values. The proposed algorithm called novel modified FPA (NMFPA) with its novel algorithmic structure offers to researchers the opportunity to: (1) use it in their comparison study (e.g., with others FPA proposed variants) and (2) develop other new methods or techniques based on its novel integrated mechanisms. To demonstrate the performance of this new FPA variant, a set of 28 benchmark functions defined in IEEE-CEC’13 and a 15 real-world numerical optimization problems proposed in the IEEE-CEC’11 are employed. Compared with FPA, two its famous variants and other state-of-the-art evolutionary algorithms, NMFPA shows overall better performance.

This paper presents an approach for the optimisation of geological disposal canister loadings, combining high resolution simulations of used nuclear fuel characteristics with an articial neural network and a genetic algorithm. The used nuclear fuels (produced in an open fuel cycle without reprocessing) considered in this work come from a Swiss Pressurised Water Reactor, taking into account their realistic lifetime in the reactor core and cooling periods, up to their disposal in the final geological repository. The case of 212 representative used nuclear fuel assemblies is analysed, assuming a loading of 4 fuel assemblies per canister, and optimizing two safety parameters: the fuel decay heat (DH) and the canister effective neutron multiplication factor k $$_{\mathrm{eff}}$$ . In the present approach, a neural network is trained as a surrogate model to evaluate the k $$_{\mathrm{eff}}$$ value to substitute the time-consuming-code Monte Carlo transport & depletion SERPENT for specific canister loading calculations. A genetic algorithm is then developed to optimise simultaneously the canister k $$_{\mathrm{eff}}$$ and DH values. The k $$_{\mathrm{eff}}$$ computed during the optimisation algorithm is using the previously developed artificial neural network. The optimisation algorithm allows (1) to minimize the number of canisters, given assumed limits for both DH and k $$_{\mathrm{eff}}$$ quantities and (2) to minimize DH and k $$_{\mathrm{eff}}$$ differences among canisters. This study represents a proof-of-principle of the neural network and genetic algorithm capabilities, and will be applied in the future to a larger number of cases.

In the majority of traditional extreme learning machine (ELM) approaches, the parameters of the basis functions are randomly generated and do not need to be tuned, while the weights connecting the hidden layer to the output layer are analytically estimated. The determination of the optimal number of basis functions to be included in the hidden layer is still an open problem. Cross-validation and heuristic approaches (constructive and destructive) are some of the methodologies used to perform this task. Recently, a deterministic algorithm based on the principal component analysis (PCA) and ELM has been proposed to assess the number of basis functions according to the number of principal components necessary to explain the 90 % of the variance in the data. In this work, the PCA part of the PCA–ELM algorithm is joined to the linear discriminant analysis (LDA) as a hybrid means to perform the pruning of the hidden nodes. This is justified by the fact that the LDA approach is outperforming the PCA one on a set of problems. Hence, the idea of combining the two approaches in a LDA–PCA–ELM algorithm is shown to be in average better than its PCA–ELM and LDA–ELM counterparts. Moreover, the performance in classification and the number of basis functions selected by the algorithm, on a set of benchmark problems, have been compared and validated in the experimental section using nonparametric tests against a set of existing ELM techniques.

The free overfall is a simple and widely used device for measuring discharge in open irrigation channels and agricultural research projects. However, the direct measurement of discharge can be difficult and time-consuming with care needed to minimize potential inaccuracies of empirical equations applied to site-specific conditions. Thus, in the present study four standalone algorithms of Isotonic Regression (ISO), Least Median of Square Regression (LMS), M5Prime (M5P) and REPT and four novel hybrid algorithms of rotation forest (ROF) combined with those four standalone models (i.e., ROF-ISO, ROF-LMS, ROF-M5P and ROF-REPT) were applied for the intelligent prediction of discharge per unit width for the free overfall condition in rectangular channels. This was accomplished via six data sets (355 data) collected from the published literature including end depth, Manning's roughness coefficient, channel width, bed slope and unit discharge. The dataset was partitioned in a 70:30 ratio randomly, 70% (248 data) of data used for model development while 30% (107 data) applied for model validation. Also, four different input combinations were constructed to identify the most effective prediction method. Furthermore, results were validated using several visually based (line graph, scatter plot, violin plot and Taylor diagram) and quantitative-based [root mean square error (RMSE), Nash–Sutcliffe efficiency (NSE), Willmott’s index of agreement, Legates and McCabe coefficient of efficiency (LM)] approaches. Results of the sensitivity analysis revealed that end depth had the highest effect on the results, while channel width was least influential. Results also showed that the best input combination incorporated all four input parameters. According to the results, ROF-REPT had the best performance with RMSE of 0.0035 (m3/s/m), NSE of 0.990, WI of 0.997% and LM of 0.905% followed by ROF-M5P REPT, M5P, ROF-LMS, ISO and LMS.

This paper reviews the potential benefits that can be obtained by the implementation of data fusion in a multi-sensor environment. A thorough review of the commonly used data fusion frameworks is presented together with important factors that need to be considered during the development of an effective data fusion problem-solving strategy. A system-based approach is defined for the application of data fusion systems within engineering. Structured guidelines for users are proposed.

In this paper, the spatial-temporal dynamics of a delayed reaction-diffusion Hopfield’s neural network model (HNNM) under a Neumann boundary condition is considered. Our main concern is the absolute, also called the delay-independent, global or local stability of the trivial and nontrivial steady states of the HNNM with non-monotone activation functions. The dissipativity of the semiflow generated by HNNM is first proved and then the absolute attractivity of both the trivial and the nontrivial steady states is obtained by adopting the idea of connecting the spatial-temporal dynamics of the HNNM with the asymptotic behaviors of a finite-dimensional discrete dynamical system. Specifically, it is shown that the strong attractors of the finite-dimensional discrete dynamical system generated by the nonlinear activation function of the HNNM are indeed the attractors of the corresponding delayed reaction-diffusion HNNM. Numerical simulations are also conducted at last to illustrate the effectiveness of our established results.

Predicting future behavior of chaotic time series and systems is a challenging area in the literature of nonlinear systems. The prediction accuracy of chaotic time series is extremely dependent on the model and the learning algorithm. In addition, the generalization property of the proposed models trained by limited observations is of great importance. In the past two decades, singular or descriptor systems and related fuzzy descriptor models have been the subjects of interest due to their many practical applications in modeling complex phenomena. In this study fuzzy descriptor models, as a more recent neurofuzzy realization of locally linear descriptor systems, which have led to the introduction of intuitive incremental learning algorithm that is called Generalized Locally Linear Model Tree algorithm, are implemented in their optimal structure to be compared with several other methods. A simple but efficient technique, based on the error indices of multiple validation sets, is used to optimize the number of neurons as well as to prevent over fitting in the incremental learning algorithms. The aim of the paper is to demonstrate the advantages of fuzzy descriptor models and to make a fair comparison between the most successful neural and neurofuzzy approaches in their best structures according to prediction accuracy, generalization, and computational complexity. The Mackey–Glass time series, Lorenz time series (as two well-known classic benchmarks), Darwin sea level pressure time series and long-term prediction of Disturbance Storm Time index, an important index of geomagnetic activity (as two natural chaotic dynamics) are used as practical examples to evaluate the power of the proposed method in long term prediction of chaotic dynamics.