Search NASA⌕ Search

SEARCH · Search NASA

Results for “Vectorized algorithm”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 883 records · Page 49

A New Hybrid Quantum-Classical Algorithm for Solving the Unit Commitment Problem

Solving problems related to planning and operations of large-scale power systems is challenging on classical computers due to their inherent nature as mixed-integer and nonlinear problems. Quantum computing provides new avenues to approach these problems. We develop a hybrid quantum-classical algorithm for the Unit Commitment (UC) problem in power systems which aims at minimizing the total cost while optimally allocating generating units to meet the hourly demand of the power loads. The hybrid algorithm combines a variational quantum algorithm (VQA) with a classical Benders-type heuristic. The resulting algorithm computes approximate solutions to UC in three stages: i) a collection of UC vectors capable meeting the power demand with lowest possible operating costs is generated based on VQA; ii) a classical sequential least squares programming (SLSQP) routine is leveraged to find the optimal power level corresponding to a predetermined number of candidate vectors; iii) in the last stage, the approximate solution of UC along with generating units power level combination is given. To demonstrate the effectiveness of the presented method, three different systems with 3 generating units, 10 generating units, and 26 generating units were tested for different time periods. In addition, convergence of the hybrid quantum-classical algorithm for select time periods is proven out on IonQ's Forte system.

Aboumrad, Willie [IonQ, Inc]↗

Algebraic grid generation using tensor product B-splines

Finite difference methods are more successful if the accompanying grid has lines which are smooth and nearly orthogonal. The development of an algorithm which produces such a grid when given the boundary description. Topological considerations in structuring the grid generation mapping are discussed. The concept of the degree of a mapping and how it can be used to determine what requirements are necessary if a mapping is to produce a suitable grid is examined. The grid generation algorithm uses a mapping composed of bicubic B-splines. Boundary coefficients are chosen so that the splines produce Schoenberg's variation diminishing spline approximation to the boundary. Interior coefficients are initially chosen to give a variation diminishing approximation to the transfinite bilinear interpolant of the function mapping the boundary of the unit square onto the boundary grid. The practicality of optimizing the grid by minimizing a functional involving the Jacobian of the grid generation mapping at each interior grid point and the dot product of vectors tangent to the grid lines is investigated. Grids generated by using the algorithm are presented.

Saunders, B. V.↗

Fast and Accurate Intersections on a Sphere

We introduce a fast, high-precision algorithm for calculating intersections between great circle arcs and lines of constant latitude on the unit sphere. We first propose a simplified intersection point formula with improved speed and numerical robustness over the ones traditionally implemented in geoscience software. We then show how algorithms based on the concept of error-free transformations (EFT) can be applied to evaluate this formula within a relative error bound that is on the order of machine precision. Here, we demonstrate that, with a vectorized and parallelized implementation, this enhanced accuracy is achieved with no compute time overhead compared to a direct calculation in hardware floating point, making our algorithm suitable for performance-sensitive applications like regridding of high-resolution climate data. In contrast, evaluating our formula using high-precision data types like quadruple precision and arbitrary precision, or using the robust intersection computation routines from the Computational Geometry Algorithms Library, leads to significant computational overhead, especially since these alternatives inhibit vectorization. More generally, our work demonstrates how EFT techniques can be combined and extended to implement nontrivial geometric calculations with high accuracy and speed.

Environmental sciences↗

Development and Validation of a Polarimetric-MCScene 3D Atmospheric Radiation Model

Polarimetric measurements can substantially enhance the ability of both spectrally resolved and single band imagery to detect the proliferation of weapons of mass destruction, providing data for locating and identifying facilities, materials, and processes of undeclared and proliferant nuclear weapons programs worldwide. Unfortunately, models do not exist that efficiently and accurately predict spectral polarized signatures for the materials of interest embedded in complex 3D environments. Having such a model would enable one to test hypotheses and optimize both the enhancement of scene contrast and the signal processing for spectral signature extraction. The Phase I set the groundwork for development of fully validated polarimetric spectral signature and scene simulation models. This has been accomplished 1. by (a) identifying and downloading state-of-the-art surface and atmospheric polarimetric data sources, (b) implementing tools for generating custom polarimetric data, and (c) identifying and requesting US Government funded field measurement data for use in validation; 2. by formulating an approach for upgrading the radiometric spectral signature model MODTRAN to generate polarimetric intensities through (a) ingestion of the polarimetric data, (b) polarimetric vectorization of existing MODTRAN modules, and (c) integration of a newly developed algorithm for computing polarimetric multiple scattering contributions; 3. by generating an initial polarimetric model that demonstrates calculation of polarimetric solar and lunar single scatter intensities arising from the interaction of incoming irradiances with molecules and aerosols; 4. by developing a design and implementation plan to (a) automate polarimetric scene construction and (b) efficiently sample polarimetric scattering and reflection events, for use in a to be developed polarimetric version of the existing first-principles synthetic scene simulation model, MCScene; and 5. by planning a validation field measurement program in collaboration with the Remote Sensing and Exploitation group at Sandia National Laboratories (SNL) in which data from their ongoing polarimetric field and laboratory measurement program will be shared and, to the extent allowed, tailored for model validation in exchange for model predictions under conditions and for geometries outside of their measurement domain.

98 NUCLEAR DISARMAMENT, SAFEGUARDS, AND PHYSICAL P↗

Machine Learning Classification of Molten Salt Heat Exchanger Channel Plugging using Synthetic Data

This report addresses the requirements of Milestone M3.4 AI capability to identify and predict maintenance events. Development of digital twins (DT) for molten salt reactor (MSR) components is crucial for reducing operating and maintenance costs (O&M) and ensuring commercial viability of these reactors. Our focus is on development of DT for MSR primary system heat exchanger (HX), a critical component, the fault in which can reduce operating efficiency and force reactor shutdown. We are investigating the feasibility of a conceptual DT of HX consisting of internal distributed temperature sensing with fiber optics and machine learning (ML) algorithms to detect and localize faults. To determine the optimal approach to detection and localization of channel plugging, we benchmark seven different ML models: Logistic Regression, K-Nearest Neighbors (KNN), Gaussian Naïve Bayes, Support Vector Machines (SVM), Decision Tree Classifier, Random Forest Tree Classifier, and Feed-Forward Neural Network. ML algorithms are benchmarked using synthetic HX plugging data generated with computational fluid dynamics COMSOL software, with added brown noise to represent experimental noise. We show that the best performance is obtained with the Decision Tree classifier.

22 GENERAL STUDIES OF NUCLEAR REACTORS↗

Algorithms for l2 and l-infinity transfer function curve fitting

In this paper algorithms for fitting transfer functions to frequency response data are developed. Given a complex vector representing the measured frequency response of a physical system, a transfer function of specified order is determined that minimizes either of the following criteria: (1) the sum of the magnitude-squared of the frequency response errors, and (2) the magnitude of the maximum error. Both of these criteria are nonlinear in the coefficients of the unknown transfer function, and iterative minimization algorithms are proposed. A numerical example demonstrates the effectiveness of the proposed algorithms.

Spanos, John T.↗

GOTTCHA Database, Version 1

One major challenge in the field of shotgun metagenomics is the accurate identification of the organisms present within the community, based on classification of short sequence reads. Though microbial community profiling methods have emerged to attempt to rapidly classify the millions of reads output from contemporary sequencers, the combination of incomplete databases, similarity among otherwise divergent genomes, and the large volumes of sequencing data required for metagenome sequencing has led to unacceptably high false discovery rates (FDR). Here we present the application of a novel, gene-independent and signature-based metagenomic taxonomic profiling tool with significantly smaller FDR, which is also capable of classifying never-before seen genomes into the appropriate parent taxa.The algorithm is based upon three primary computational phases: (I) genomic decomposition into bit vectors, (II) bit vector intersections to identify shared regions, and (III) bit vector subtractions to remove shared regions and reveal unique, signature regions.In the Decomposition phase, genomic data is first masked to highlight only the valid (non-ambiguous) regions and then decomposed into overlapping 24-mers. The k-mers are sorted along with their start positions, de-replicated, and then prefixed, to minimize data duplication. The prefixes are indexed and an identical data structure is created for the start positions to mimic that of the k-mer data structure.During the Intersection phase -- which is the most computationally intensive phase -- as an all-vs-all comparison is made, the number of comparisons is first reduced by four methods: (a) Prefix restriction, (b) Overlap detection, (c) Overlap restriction, and (d) Result recording. In Prefix restriction, only k-mers of the same prefix are compared. Within that group, potential overlap of k-mer suffixes that would result in a non-empty set intersection are screened for. If such an overlap exists, the region which intersects is first reduced by performing a binary search of the boundary suffixes of the smaller set into the larger set, which defines the limits of the zipper-based intersection process. Rather than recording the actual k-mers of the intersection, another data structure of identical "shape" is created which consists of only bit vectors so that only a 1 or 0 will be stored in the location of the k-mer suffix that was found in the intersection. This reduces the amount of data generated and stored considerably.During the Subtraction phase, relevant intersection bitmasks are first unionized together to form a single bitmask which is then applied over the original genome to reveal only those regions of the genome that are unique. These regions are then exported to disk in FASTA format and used in the application of determining the constituents of an unknown metagenomic community.The DATABASE provided is the result of the algorithm described.

Freitas, Tracey↗

GOTTCHA Database, Version 1

One major challenge in the field of shotgun metagenomics is the accurate identification of the organisms present within the community, based on classification of short sequence reads. Though microbial community profiling methods have emerged to attempt to rapidly classify the millions of reads output from contemporary sequencers, the combination of incomplete databases, similarity among otherwise divergent genomes, and the large volumes of sequencing data required for metagenome sequencing has led to unacceptably high false discovery rates (FDR). Here we present the application of a novel, gene-independent and signature-based metagenomic taxonomic profiling tool with significantly smaller FDR, which is also capable of classifying never-before seen genomes into the appropriate parent taxa.The algorithm is based upon three primary computational phases: (I) genomic decomposition into bit vectors, (II) bit vector intersections to identify shared regions, and (III) bit vector subtractions to remove shared regions and reveal unique, signature regions.In the Decomposition phase, genomic data is first masked to highlight only the valid (non-ambiguous) regions and then decomposed into overlapping 24-mers. The k-mers are sorted along with their start positions, de-replicated, and then prefixed, to minimize data duplication. The prefixes are indexed and an identical data structure is created for the start positions to mimic that of the k-mer data structure.During the Intersection phase -- which is the most computationally intensive phase -- as an all-vs-all comparison is made, the number of comparisons is first reduced by four methods: (a) Prefix restriction, (b) Overlap detection, (c) Overlap restriction, and (d) Result recording. In Prefix restriction, only k-mers of the same prefix are compared. Within that group, potential overlap of k-mer suffixes that would result in a non-empty set intersection are screened for. If such an overlap exists, the region which intersects is first reduced by performing a binary search of the boundary suffixes of the smaller set into the larger set, which defines the limits of the zipper-based intersection process. Rather than recording the actual k-mers of the intersection, another data structure of identical "shape" is created which consists of only bit vectors so that only a 1 or 0 will be stored in the location of the k-mer suffix that was found in the intersection. This reduces the amount of data generated and stored considerably.During the Subtraction phase, relevant intersection bitmasks are first unionized together to form a single bitmask which is then applied over the original genome to reveal only those regions of the genome that are unique. These regions are then exported to disk in FASTA format and used in the application of determining the constituents of an unknown metagenomic community.The DATABASE provided is the result of the algorithm described.

Freitas, Tracey↗

SU-F-BRF-04: Evaluation of Five Commercially-Available Algorithms for Deformable Image Registration

Purpose: Deformable image registration (DIR) is increasingly being used in various clinical applications. Although there are several DIR packages all making successful attempts at modeling complex anatomical changes using even more complex mathematical approximations, they are all subject to various uncertainties. Many studies have attempted to quantify the spatial uncertainty with DIR. This is the first study to compare the uncertainty for interfraction DIR for 5 different commercially-available algorithms. The aim of this study was to benchmark the performance of the most commonlyused DIR algorithms offered through these 5 software packages: Eclipse, MIM, Pinnacle, RaySearch, and Velocity. Methods: A set of 10 virtual H'N phantoms [Pukala et al. MedPhys. 40(11) 2013] with known deformations were used to determine the spatial errors that might be seen when performing DIR. The “ground-truth” deformation vector field (DVF) was compared to the DVF output of the 5 commercially-available algorithms in order to evaluate spatial errors for six regions of interest (ROIs): brainstem, cord, mandible, left parotid, right parotid, and the external body contour. Results: We found that each software package had varying uncertainties with the various ROIs, but were generally all comparable to one another – with mean spatial errors for each algorithm below 3.5 mm for each ROI (averaged across all phantoms). We also found that no single algorithm was the clear winner over the other 4 algorithms. However, at times, we found huge maximum errors in our results (e.g. phantom #9 maximum errors: right parotid = 22.9 mm, external contour = 30.5mm) with the varying DIR algorithms. Conclusion: Although our evaluation was limited to H'N patients, we show that our methods are a single-assessment analysis tool that could be used by any physicist, within any type of facility, to compare their DIR software before initiating widespread use within their daily radiotherapy practice.

60 APPLIED LIFE SCIENCES↗

Trinity Benchmarks on the Intel Xeon Phi (Knights Corner)

This report documents the early experiences with porting and performance analysis of the Tri-Lab Trinity benchmark applications on Intel Xeon Phi (Knights Corner) (KNC) processor. KNC, the second generation of the Intel Many Integrated Core (MIC) architectures, uses a large number of small P54C-x86 cores with wide vector units and is deployed as PCI bus attached process accelerators. Sandia has experimental test beds of small InifiniBand clusters and workstations to investigate the performance of the MIC architecture. On these experimental test beds the programming models that may be investigated are "offload", "symmetric" and "native". Among these program usage models our primary interest is in the so called "native" mode, because the planned Trinity system to be deployed in 2016 using the next generation MIC processor architecture called Knights Landing would be self-hosted. Trinity / NERSC-8 benchmark programs cover a variety of scientific disciplines and they were used to guide the procurement of these systems. Architectures such as the Intel MIC are well suited to study evolving processor architectures and a usage model commonly referred to as MPI + X that facilitates migration of our applications to use both coarse grain and fine grain parallelism. Our focus with the applications selected is on the efficacy of algorithms in these applications to take advantage of features like: large number of cores, wide vector units, higher-bandwidth and deeper memory sub-system. This is a first step towards understanding applications, algorithms and programming environments for Trinity and future exascale computing systems.

97 MATHEMATICS AND COMPUTING↗

Parallel CFD Algorithms for Aerodynamical Flow Solvers on Unstructured Meshes

The Advisory Group for Aerospace Research and Development (AGARD) has requested my participation in the lecture series entitled Parallel Computing in Computational Fluid Dynamics to be held at the von Karman Institute in Brussels, Belgium on May 15-19, 1995. In addition, a request has been made from the US Coordinator for AGARD at the Pentagon for NASA Ames to hold a repetition of the lecture series on October 16-20, 1995. I have been asked to be a local coordinator for the Ames event. All AGARD lecture series events have attendance limited to NATO allied countries. A brief of the lecture series is provided in the attached enclosure. Specifically, I have been asked to give two lectures of approximately 75 minutes each on the subject of parallel solution techniques for the fluid flow equations on unstructured meshes. The title of my lectures is "Parallel CFD Algorithms for Aerodynamical Flow Solvers on Unstructured Meshes" (Parts I-II). The contents of these lectures will be largely review in nature and will draw upon previously published work in this area. Topics of my lectures will include: (1) Mesh partitioning algorithms. Recursive techniques based on coordinate bisection, Cuthill-McKee level structures, and spectral bisection. (2) Newton's method for large scale CFD problems. Size and complexity estimates for Newton's method, modifications for insuring global convergence. (3) Techniques for constructing the Jacobian matrix. Analytic and numerical techniques for Jacobian matrix-vector products, constructing the transposed matrix, extensions to optimization and homotopy theories. (4) Iterative solution algorithms. Practical experience with GIVIRES and BICG-STAB matrix solvers. (5) Parallel matrix preconditioning. Incomplete Lower-Upper (ILU) factorization, domain-decomposed ILU, approximate Schur complement strategies.

Barth, Timothy J.↗

Automated Conflict Resolution For Air Traffic Control

The ability to detect and resolve conflicts automatically is considered to be an essential requirement for the next generation air traffic control system. While systems for automated conflict detection have been used operationally by controllers for more than 20 years, automated resolution systems have so far not reached the level of maturity required for operational deployment. Analytical models and algorithms for automated resolution have been traffic conditions to demonstrate that they can handle the complete spectrum of conflict situations encountered in actual operations. The resolution algorithm described in this paper was formulated to meet the performance requirements of the Automated Airspace Concept (AAC). The AAC, which was described in a recent paper [1], is a candidate for the next generation air traffic control system. The AAC's performance objectives are to increase safety and airspace capacity and to accommodate user preferences in flight operations to the greatest extent possible. In the AAC, resolution trajectories are generated by an automation system on the ground and sent to the aircraft autonomously via data link .The algorithm generating the trajectories must take into account the performance characteristics of the aircraft, the route structure of the airway system, and be capable of resolving all types of conflicts for properly equipped aircraft without requiring supervision and approval by a controller. Furthermore, the resolution trajectories should be compatible with the clearances, vectors and flight plan amendments that controllers customarily issue to pilots in resolving conflicts. The algorithm described herein, although formulated specifically to meet the needs of the AAC, provides a generic engine for resolving conflicts. Thus, it can be incorporated into any operational concept that requires a method for automated resolution, including concepts for autonomous air to air resolution.

Erzberger, Heinz↗

NASA Langley's Formal Methods Research in Support of the Next Generation Air Transportation System

This talk will provide a brief introduction to the formal methods developed at NASA Langley and the National Institute for Aerospace (NIA) for air traffic management applications. NASA Langley's formal methods research supports the Interagency Joint Planning and Development Office (JPDO) effort to define and develop the 2025 Next Generation Air Transportation System (NGATS). The JPDO was created by the passage of the Vision 100 Century of Aviation Reauthorization Act in Dec 2003. The NGATS vision calls for a major transformation of the nation s air transportation system that will enable growth to 3 times the traffic of the current system. The transformation will require an unprecedented level of safety-critical automation used in complex procedural operations based on 4-dimensional (4D) trajectories that enable dynamic reconfiguration of airspace scalable to geographic and temporal demand. The goal of our formal methods research is to provide verification methods that can be used to insure the safety of the NGATS system. Our work has focused on the safety assessment of concepts of operation and fundamental algorithms for conflict detection and resolution (CD&R) and self- spacing in the terminal area. Formal analysis of a concept of operations is a novel area of application of formal methods. Here one must establish that a system concept involving aircraft, pilots, and ground resources is safe. The formal analysis of algorithms is a more traditional endeavor. However, the formal analysis of ATM algorithms involves reasoning about the interaction of algorithmic logic and aircraft trajectories defined over an airspace. These trajectories are described using 2D and 3D vectors and are often constrained by trigonometric relations. Thus, in many cases it has been necessary to unload the full power of an advanced theorem prover. The verification challenge is to establish that the safety-critical algorithms produce valid solutions that are guaranteed to maintain separation under all possible scenarios. Current research has assumed perfect knowledge of the location of other aircraft in the vicinity so absolute guarantees are possible, but increasingly we are relaxing the assumptions to allow incomplete, inaccurate, and/or faulty information from communication sources.

Butler, Ricky W.↗

Euler/Navier-Stokes calculations of transonic flow past fixed- and rotary-wing aircraft configurations

Computational fluid dynamics has an increasingly important role in the design and analysis of aircraft as computer hardware becomes faster and algorithms become more efficient. Progress is being made in two directions: more complex and realistic configurations are being treated and algorithms based on higher approximations to the complete Navier-Stokes equations are being developed. The literature indicates that linear panel methods can model detailed, realistic aircraft geometries in flow regimes where this approximation is valid. As algorithms including higher approximations to the Navier-Stokes equations are developed, computer resource requirements increase rapidly. Generation of suitable grids become more difficult and the number of grid points required to resolve flow features of interest increases. Recently, the development of large vector computers has enabled researchers to attempt more complex geometries with Euler and Navier-Stokes algorithms. The results of calculations for transonic flow about a typical transport and fighter wing-body configuration using thin layer Navier-Stokes equations are described along with flow about helicopter rotor blades using both Euler/Navier-Stokes equations.

Deese, J. E.↗

Validation of deformable image registration algorithms on CT images of ex vivo porcine bladders with fiducial markers

Purpose: The spatial accuracy of deformable image registration (DIR) is important in the implementation of image guided adaptive radiotherapy techniques for cancer in the pelvic region. Validation of algorithms is best performed on phantoms with fiducial markers undergoing controlled large deformations. Excised porcine bladders, exhibiting similar filling and voiding behavior as human bladders, provide such an environment. The aim of this study was to determine the spatial accuracy of different DIR algorithms on CT images ofex vivo porcine bladders with radiopaque fiducial markers applied to the outer surface, for a range of bladder volumes, using various accuracy metrics. Methods: Five excised porcine bladders with a grid of 30–40 radiopaque fiducial markers attached to the outer wall were suspended inside a water-filled phantom. The bladder was filled with a controlled amount of water with added contrast medium for a range of filling volumes (100–400 ml in steps of 50 ml) using a luer lock syringe, and CT scans were acquired at each filling volume. DIR was performed for each data set, with the 100 ml bladder as the reference image. Six intensity-based algorithms (optical flow or demons-based) implemented in theMATLAB platform DIRART, a b-spline algorithm implemented in the commercial software package VelocityAI, and a structure-based algorithm (Symmetric Thin Plate Spline Robust Point Matching) were validated, using adequate parameter settings according to values previously published. The resulting deformation vector field from each registration was applied to the contoured bladder structures and to the marker coordinates for spatial error calculation. The quality of the algorithms was assessed by comparing the different error metrics across the different algorithms, and by comparing the effect of deformation magnitude (bladder volume difference) per algorithm, using the Independent Samples Kruskal-Wallis test. Results: The authors found good structure accuracy without dependency on bladder volume difference for all but one algorithm, and with the best result for the structure-based algorithm. Spatial accuracy as assessed from marker errors was disappointing for all algorithms, especially for large volume differences, implying that the deformations described by the registration did not represent anatomically correct deformations. The structure-based algorithm performed the best in terms of marker error for the large volume difference (100–400 ml). In general, for the small volume difference (100–150 ml) the algorithms performed relatively similarly. The structure-based algorithm exhibited the best balance in performance between small and large volume differences, and among the intensity-based algorithms, the algorithm implemented in VelocityAI exhibited the best balance. Conclusions: Validation of multiple DIR algorithms on a novel physiological bladder phantom revealed that the structure accuracy was good for most algorithms, but that the spatial accuracy as assessed from markers was low for all algorithms, especially for large deformations. Hence, many of the available algorithms exhibit sufficient accuracy for contour propagation purposes, but possibly not for accurate dose accumulation.

60 APPLIED LIFE SCIENCES↗

Vectorization techniques for probability distribution functions using VecCore

Probability distribution functions (PDFs) are very used in modeling random processes and physics simulations. Improving the performance of algorithms that generate many random numbers under complex PDFs is often a very challenging task when methods as direct functions are not available. In this work we present general strategies on how to vectorize some PDFs using VecCore library. We show the results for the Exponential, Gaussian, discrete Poisson and Gamma probability distributions.

Chaparro Amaro, Oscar R.↗

Experimental demonstration of coupled multi-peak Bragg coherent diffraction imaging with genetic algorithms

Bragg coherent diffraction imaging has the potential to provide significant insight into the structure-properties relationship for crystalline materials by imaging, with nanoscale resolution, three-dimensional strain fields within individual grains and nanoparticles. The capability of present-day synchrotrons to locate and measure a multiplicity of Bragg reflections from a single grain makes it possible to recover the full strain tensor with nanometer resolution. Recent methods for coupling reconstructions from several peaks to determine the strain tensor have been developed and applied to synthetic data, but have not been applied to experimental data. Here, using a coupled genetic reconstruction algorithm, we reconstruct an experimental data set and demonstrate improvements in the ability to resolve vector-valued displacement fields internal to the particle as compared to what is achieved with a noncoupled approach. The coupled approach developed in this work was also validated on simulated data sets. In both simulated and experimental data, reconstructions from our coupled Bragg peak algorithm show improvements over the noncoupled independent reconstruction method of 5% in terms of accuracy and 53% in terms of consistency.

75 CONDENSED MATTER PHYSICS, SUPERCONDUCTIVITY AND↗

Accelerated Training for Large Feedforward Neural Networks

In this paper we introduce a new training algorithm, the scaled variable metric (SVM) method. Our approach attempts to increase the convergence rate of the modified variable metric method. It is also combined with the RBackprop algorithm, which computes the product of the matrix of second derivatives (Hessian) with an arbitrary vector. The RBackprop method allows us to avoid computationally expensive, direct line searches. In addition, it can be utilized in the new, 'predictive' updating technique of the inverse Hessian approximation. We have used directional slope testing to adjust the step size and found that this strategy works exceptionally well in conjunction with the Rbackprop algorithm. Some supplementary, but nevertheless important enhancements to the basic training scheme such as improved setting of a scaling factor for the variable metric update and computationally more efficient procedure for updating the inverse Hessian approximation are presented as well. We summarize by comparing the SVM method with four first- and second- order optimization algorithms including a very effective implementation of the Levenberg-Marquardt method. Our tests indicate promising computational speed gains of the new training technique, particularly for large feedforward networks, i.e., for problems where the training process may be the most laborious.

Stepniewski, Slawomir W.↗