Turco A.,ESTECO Srl
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) | Year: 2010
AFSQP is a Sequential Quadratic Programming algorithm which obtains global convergence through an adaptive filter technique. This adaptivity is the major innovation in this work. The resulting algorithm can deal with constraints involving different length scales without requiring their normalization. The effort related to gradients computation is compensated by achieving superlinear local convergence rate (under some hypothesis on the problem, the algorithm can reach quadratic rates). Second order derivatives are approximated with classical BFGS formula and need not to be computed. We describe the theoretical background of the algorithm as well as its implementation details. A comparison between AFSQP and four different SQP implementations is performed considering several small and medium scale problems selected within Hoch and Schittkowski suite. We focus attention on the number of point evaluations required. © 2010 Springer-Verlag.
Turco A.,ESTECO Srl
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) | Year: 2011
We propose a metamodel approach to the approximation of functions gradients within a hybrid genetic algorithm. The underlying structure is implemented in order to support parallel execution of the code: a genetic and a SQP algorithm run in different threads and can ask designs evaluations independently, but keeping all the available resources always working. A common archive collects the results and generates the population for the GA and the starting points for the SQP runs. A particular attention is dedicated to elitism and to constraints. The hybridization is performed through a modified e - constrained method. The general philosophy of the algorithm is to concentrate on not wasting information: metamodels, archiving and elitism, steady-state parallel evolution are key elements for this scope and they will be discussed in details. A preliminary but explanatory row of tests concludes the paper highlighting the benefits of this new approach. © Springer-Verlag Berlin Heidelberg 2011.
Xue Z.,ESTECO North America |
Marchi M.,ESTECO Srl |
Parashar S.,ESTECO North America |
Li G.,Ford Motor Company
SAE Technical Papers | Year: 2015
Robustness/Reliability Assessment and Optimization (RRAO) is often computationally expensive because obtaining accurate Uncertainty Quantification (UQ) may require a large number of design samples. This is especially true where computationally expensive high fidelity CAE simulations are involved. Approximation methods such as the Polynomial Chaos Expansion (PCE) and other Response Surface Methods (RSM) have been used to reduce the number of time-consuming design samples needed. However, for certain types of problems require the RRAO, one of the first question to consider is which method can provide an accurate and affordable UQ for a given problem. To answer the question, this paper tests the PCE, RSM and pure sampling based approaches on each of the three selected test problems: the Ursem Waves mathematical function, an automotive muffler optimization problem, and a vehicle restraint system optimization problem. Results of the UQ are compared thoroughly and recommendations based on the empirical results are made as the design guidelines to engineers. Copyright © 2015 SAE International.
Costanzo S.,ESTECO Srl |
Costanzo S.,University of Trieste |
Castelli L.,University of Trieste |
Turco A.,ESTECO Srl
Engineering Optimization IV - Proceedings of the 4th International Conference on Engineering Optimization, ENGOPT 2014 | Year: 2014
Genetic algorithms are versatile tools that are able to tackle a wide range of real-world problems. In this paper, we propose a general-purpose genetic algorithm for black box multi-objective optimization well suited for different types of variables (continuous, discrete, and combinatorial) and constraints (linear and nonlinear, equalities and inequalities) without requiring any customization, such as problem-dependent operators. The basic idea is to extract as much information as possible from the characteristics of the decision variables in order to activate the most appropriate routines automatically. We apply this strategy to a real world problem: The layout optimization of a Wireless Sensor Network (WSN). This problem can be equivalently formulated in two different ways, both presenting some critical points for an effective application of standard genetic algorithms. We show how our algorithm can learn from its structure and solve the problem more efficiently than other classical genetic approaches. © 2015 Taylor & Francis Group, London.
Turco A.,ESTECO Srl |
Kavka C.,ESTECO Srl
International Journal of Innovative Computing and Applications | Year: 2011
We present a multi-objective genetic algorithm called magnifying front genetic algorithm (MFGA) designed in order to treat complex real-world optimisation problems. A first source of complexity is the presence of different input variables classes (real, discrete and categorical). MFGA is able to treat appropriately each of them as well as any combination. Moreover, real-world applications often require a long time to evaluate objective values from input variables. We deal with this issue working on elitism (in order to tune properly the balance between explorative and exploitative capabilities of the algorithm) and introducing a parallel steady-state evolution scheme, which is able to use the available computing resources as much intensively as possible. We test the algorithm on two different scenarios: mathematical benchmarks and real-world applications. For the latter one we chose a problem arising in multi-processor system-on-chip (MPSoC) design, a field which is characterised by discrete and more often categorical variables. © 2011 Inderscience Enterprises Ltd.