Danilo R. B. Araújo

A performance comparison of multi-objective optimization evolutionary algorithms for all-optical networks design

In this paper we investigate the performance of well known multi-objective optimization evolutionary algorithms (MOEA) applied to the design of all-optical networks. We focused on the simultaneous optimization of the network topology and the device specifications in order to both minimize the total cost to build the network, i.e. the capital expenditure, and to maximize the overall network performance. We used the network blocking probability to assess the quality of the network service. We have considered the following five different MOEA: NSGAII, SPEA2, PESAII, PAES and MODE. In order to suggest a suitable algorithm to solve the problem, we performed a set of simulations aiming to analyze the convergence ability and the diversity of the generated solutions. We used four well known metrics to compare the achieved Pareto Fronts: hypervolume, spacing, maximum spread and coverage. From our results, we believe that the NSGAII and the SPEA2 algorithms are more suitable to solve this specific problem.

An evolutionary approach with surrogate models and network science concepts to design optical networks

Physical topology design of optical networks is frequently accomplished by using evolutionary approaches. However, fitness evaluation for this type of problems is time consuming and the overall optimization process presents a huge execution time. In this paper we propose a new method that uses a multi-objective evolutionary approach to handle the design of all-optical networks. We focused on the simultaneous optimization of the network topology and the device specifications in order to minimize both the capital expenditure of the network and the network performance. Our method uses surrogate models to accelerate the fitness evaluation and a novel network generative model based on preferential attachment to generate the seeds for the evolutionary process. Our approach can provide high quality solutions with a very small execution time when compared to the previous approaches. In order to assess our proposal we performed a set of simulations aiming to analyze the convergence ability and the diversity of the generated solutions for scenarios considering uniform and non-uniform traffic matrices. From our results, we obtained an evolutionary approach that presents better solutions than previous proposals for all analyzed scenarios. Our proposal presents an execution time that is up to 84% and 88% lower than the execution time needed by the previous approaches for uniform and non-uniform traffic, respectively.

Methodology to obtain a fast and accurate estimator for blocking probability of optical networks

The assessment of optical networks considering physical impairments is frequently accomplished by using time-consuming analysis tools.We propose in this paper to use artificial neural networks to predict the blocking probability of optical networks with dynamic traffic by using topological metrics and general information of the physical layer. The training process is accomplished by supervised learning based on a historical database of networks. We also propose a new and simple topological property to represent the capacity of the network to distribute traffic. From the results, we found that this novel topological property improves the estimator accuracy. We compared the results of our proposal with the outcome of a discrete event simulator for optical networks. The simulator provides an estimate for blocking probability of alloptical networks considering physical impairments. We show that our approach is faster than discrete event simulators; we obtained a speedup of greater than 7500 times, with comparable estimation errors.

Design of transparent optical networks considering physical impairments, CAPEX and energy consumption

We propose a methodology which applies a multi-objective Evolutionary Algorithm, the NSGAII, in order to design transparent optical networks, aiming to minimize simultaneously the total cost to build the network, the blocking probability and the energy consumption during operation. The optimizer provides a set of non-dominated solutions and, after that, the designer can decide which solution is more suitable for a specific case. We believe these three different important aspects must be taken into account during the design process. However, just some few papers tackle energy consumption from a physical network design perspective. Besides, to the best of our knowledge, none of the previous presented proposals consider all these issues simultaneously.

An efficient multi-objective evolutionary optimizer to design all-optical networks considering physical impairments and CAPEX

In this paper we propose efficient operators for a well known multi-objective evolutionary optimizer, called NSGA II, applied to design all-optical networks regarding the network topology and the device specifications in order to both minimize the capital expenditure to build the network and to maximize the overall network performance. From the experiments, we perceived that it is better to use an uniform crossover, to include preferences a priori and to initialize the individuals emphasizing the network topology.

Artificial Neural Networks to estimate Blocking Probability of transparent optical networks: A robustness study for different networks

Recent studies demonstrated the advantage of alternative methods to assess optical networks based on Artificial Neural Networks (ANN), which is to obtain a fast estimation of Blocking Probability (BP) with a small error. In previous works we proposed the use of ANNs to predict the BP of optical networks with dynamic traffic by using topological metrics and general information of the physical layer. In this paper we use the node locations of six deployed networks in order to evaluate the robustness of the estimator. We also propose four new measures related to physical layer and we compare the results of our proposal with the outcome of a discrete event network simulator. From our results we conclude that ANN is a promising technique to estimate the BP of transparent optical networks because we obtained a fast BP estimation with small errors for all analyzed networks.

Using the Entropy of the DFT of the Laplacian Eigenvalues to Assess Networks

There are several metrics that are very useful to analyze and to design networks. These metrics, including the spectral-based ones, can be used to retrieve topological properties from the network. We observed that if one applies the Discrete Fourier Transform (DFT) over the eigenvalues of the Laplacian matrix, it is possible to observe different patterns in the DFT depending on some properties of the analyzed networks. In this paper, we propose a new metrics based on the entropy of the DFT samples, that can be used to identify the type of network. We evaluated this metrics in networks generated by four different procedures (k-Regular, Erdos-Renyi, Watts-Strogatz and Barabasi-Albert) and in well-known datasets of real networks. The results indicate that one can use the proposed metrics to identify the generational model of the network.

New Graph Model to Design Optical Networks

The design of optical networks is frequently accomplished by using evolutionary algorithms (EAs). However, the overall optimization process presents a huge execution time. In this letter, we propose a new method based on factorial design and geographical graph models to design optical networks. We propose an iterative graph generator based on Gabriel graphs that considers the number of deployed fibers and traffic demand. We focus on the specification of the fiber topology and optical devices aiming at finding a good trade-off in terms of capital expenditure and blocking probability. Our proposal provides high quality solutions with a very small execution time when compared to EAs. From our results, our proposal spends less than 1% of the time required by EAs and achieves better results.

Using Multi-Layer Perceptron and complex network metrics to estimate the performance of optical networks

The performance assessment of a WDM network considering physical impairments is a difficult task and is frequently accomplished by using time consuming computational simulations. On the other hand, we observed that several metrics have been proposed to assess different aspects of a network structure. In this paper we propose that a set of metrics can be combined in order to obtain a fast estimation of a WDM network performance, based on a historical database of networks. The estimator was obtained by means of the most used Artificial Neural Network (ANN) architecture, called Multi-Layer Perceptron, that was trained using the classical back-propagation algorithm. According to our results, it is possible to build an estimator based on network metrics that assess WDM networks considering the trade-off between the processing time and the precision of the results. Our study also suggests that this kind of estimator can be easily adapted to other scenarios of WDM networks since Artificial Neural Networks present interesting characteristics, such as adaptation and flexibility.

Analyzing surrogate models to assess Blocking Probability of optical networks

Recent studies demonstrated the feasibility of surrogate methods to assess optical networks based on Artificial Neural Networks (ANNs). However, surrogate methods present different trade offs between accuracy and resource utilization efficiency, such as computational time. In this paper we analyze the use of ANN to forecast the Blocking Probability (BP) of deployed optical networks considering different architectures for the underlying alternative method. We also analyze the impact of the adopted physical layer model and the number of optical networks needed to train the ANN. We compare the results of our proposal with the outcome of a discrete event network simulator. From our results we can conclude that ANN is a promising technique to estimate the BP of transparent optical networks, but the dataset used to train the ANN and the physical layer model are crucial for the proper design of this type of tool.