Number Of Outer Iterations Research Articles

Joint Precoding and CSI Dimensionality Reduction: An Efficient Deep Unfolding Approach

A recently proposed unified precoding and pilot design optimization (UPPiDO) framework offers a reduction in both training and feedback overhead of acquiring channel state information (CSI) and an enhancement in robustness (to CSI uncertainties) at the expense of a more computationally demanding precoding optimization. To address this increased complexity, in this paper we first propose an unfolding-friendly iterative algorithm, which can efficiently address a family of non-convex and non-smooth problems. Then, we develop an efficient approach to unfold the iterative algorithm designed. Besides being applicable to important and typical iterative optimization algorithms, a pivotal advantage of the proposed unfolding approach is that the trainable parameters are scalars (rather than matrices). This, in turn, reduces the number of training samples required and makes it suitable for rapidly fluctuating wireless environments. We apply the algorithm unfolding (AU) techniques developed to our UPPiDO-based symbol-level precoding and block-level precoding. Our complexity analysis indicates that the computational complexity is scalable both with the numbers of served users and antennas. Our simulation results demonstrate that the number of outer iterations (or layers) required is about 1/3 of that of the original iterative algorithms.

Improved Wielandt method with real-time error tolerance based on NLSP3 code

The Wielandt method reduces the number of outer iterations but increases the number of inner iterations. The number of calculated inner iterations should be reduced to improve the efficiency of the Wielandt method, and then the error tolerance for the Wielandt inner iteration should be larger. We give a new formula to calculate the error tolerance and achieve the real-time adjustment of the error tolerance for the Wielandt inner iteration. The improved Wielandt method is implemented in NLSP3 code, and tested by the 2D-IAEA benchmark and the 3D-IAEA benchmark. The results show that the improved Wielandt method can reduce the total number of calculated inner iterations, thus improving the efficiency of the Wielandt method.

An inexact accelerated stochastic ADMM for separable convex optimization

An inexact accelerated stochastic Alternating Direction Method of Multipliers (AS-ADMM) scheme is developed for solving structured separable convex optimization problems with linear constraints. The objective function is the sum of a possibly nonsmooth convex function and a smooth function which is an average of many component convex functions. Problems having this structure often arise in machine learning and data mining applications. AS-ADMM combines the ideas of both ADMM and the stochastic gradient methods using variance reduction techniques. One of the ADMM subproblems employs a linearization technique while a similar linearization could be introduced for the other subproblem. For a specified choice of the algorithm parameters, it is shown that the objective error and the constraint violation are $$\mathcal {O}(1/k)$$ relative to the number of outer iterations k. Under a strong convexity assumption, the expected iterate error converges to zero linearly. A linearized variant of AS-ADMM and incremental sampling strategies are also discussed. Numerical experiments with both stochastic and deterministic ADMM algorithms show that AS-ADMM can be particularly effective for structured optimization arising in big data applications.

The Stability of Linear Diffusion Acceleration Relative to CMFD

Coarse Mesh Finite Difference (CMFD) is a widely-used iterative acceleration method for neutron transport problems in which nonlinear terms are introduced in the derivation of the low-order CMFD diffusion equation. These terms, including the homogenized diffusion coefficient, the current coupling coefficients, and the multiplicative prolongation constant, are subject to numerical instability when a scalar flux estimate becomes sufficiently small or negative. In this paper, we use a suite of contrived problems to demonstrate the susceptibility of CMFD to failure for each of the vulnerable quantities of interest. Our results show that if a scalar flux estimate becomes negative in any portion of phase space, for any iterate, numerical instability can occur. Specifically, the number of outer iterations required for convergence of the CMFD-accelerated transport problem can increase dramatically, or worse, the iteration scheme can diverge. An alternative Linear Diffusion Acceleration (LDA) scheme addresses these issues by explicitly avoiding local nonlinearities. Our numerical results show that the rapid convergence of LDA is unaffected by the very small or negative scalar flux estimates that can adversely affect the performance of CMFD. Therefore, our results demonstrate that LDA is a robust alternative to CMFD for certain sensitive problems in which CMFD can exhibit reduced effectiveness or failure.

Impact of Various Parameters on Convergence Performance of CMFD Acceleration for MOC in Multigroup Heterogeneous Geometry

The impact of various parameters in the coarse mesh finite difference (CMFD) acceleration method on overall convergence behavior is investigated through numerical calculations using the method of characteristics (MOC). Four parameters appearing in the CMFD acceleration with MOC, i.e., scalar flux distribution in flat flux regions (FFRFlux), the scalar flux distribution in CMFD meshes (CMFDFlux), homogenized cross sections (HXSs) in CMFD meshes, and current correction factors (CCFs), are considered. Parts of these four parameters are fixed to the converged values throughout iterations in order to estimate their impact on convergence. Numerical calculations are carried out for Korea Advanced Institute of Science and Technology’s (KAIST’s) benchmark problem KAIST-2A, which is a heterogeneous and multigroup problem, and the number of outer iterations to reach convergence is evaluated. The impact of geometric heterogeneity and cross-section homogenization in the CMFD acceleration has not been considered in linearized Fourier analysis so far. The calculation results indicate that (1) convergence of HXS has little impact on the overall convergence, (2) convergence of FFRFlux is dominant followed by CCF when a CMFD mesh is optically thin, and (3) convergence of FFRFlux is dominant when a CMFD mesh is optically thick and contains many flat flux regions.

Consistent pCMFD Acceleration Schemes of the Three-Dimensional Transport Code PROTEUS-MOC

This paper presents the new acceleration schemes implemented in the three-dimensional (3-D) transport solver PROTEUS-MOC in conjunction with the fixed-point iteration (FPI) methods based on a single generalized minimal residual (GMRES) iteration and one or two transport sweeps per group in each outer iteration. In order to adopt a FPI scheme that employs only one or two inner iterations, single- and two-level consistent partial current–based coarse-mesh finite difference (pCMFD) acceleration methods were implemented to remove the instability problem of the consistent coarse-mesh finite difference (CMFD) method encountered when the inner iteration convergence is not sufficiently tight. In the spatial two-level acceleration method to speed up the lower-order diffusion calculations, the first level solves a fine-mesh finite difference fixed-source problem and the second level solves a CMFD eigenvalue problem. The implemented acceleration schemes were tested using the C5G7 benchmark problems, a critical core configuration of the Transient Reactor Test Facility (TREAT), and a C5G7 transient benchmark problem. Numerical test results showed that the consistent pCMFD acceleration is always stable even for the FPI methods with one inner iteration and that the single transport sweep method is always more efficient than the single GMRES iteration method. It was also observed that the two-level pCMFD acceleration in conjunction with the FPI with single transport sweep per outer iteration is very effective in reducing the number of outer iterations and the lower-order diffusion calculation time. Compared to the current iteration scheme of PROTEUS-MOC with fully converged GMRES iteration without acceleration, this acceleration reduced the total computational time by factors of 33.7, 19.9, and 26.0 for the two-dimensional C5G7, 3-D C5G7, and TREAT M8CAL criticality problems, respectively. The gain was even much larger for transient fixed-source problems (TFSPs) that are near critical. The speedup factor was 100 for one TFSP with subcriticality level of 40 mk and 519 for another TFSP with subcriticality level of 9 mk.

Generalized Partitioned Matrix Acceleration for Variational Nodal Diffusion Method

The Generalized Partitioned Matrix (GPM) acceleration method for the Variational Nodal Method (VNM) with diffusion approximation is presented. In the GPM method, the vectors of expansion coefficients of the scalar flux, source, and partial currents are divided into low-order and high-order terms. For each outer iteration, the low-order terms of the flux, fission source, and partial currents are first solved with fixed higher-order terms from the preceding outer iteration, and then a full matrix sweep through the energy groups is performed to update the full set of expansion coefficients. The GPM method increases the CPU time per outer iteration but reduces the overall computational time significantly by reducing the number of outer iterations required for convergence. The GPM acceleration method has been implemented in the NODAL code, and its performance was compared with that of the traditional Partitioned Matrix (PM) acceleration scheme for four problems: two- and three-dimensional C5G7 problems, a NuScale modular core problem, and a large pressurized water reactor problem. The numerical results show that the PM acceleration consistently reduces the computational time by a factor of 2.0 and the GPM acceleration yields two to three times higher speedup than with PM acceleration by reducing the number of outer iterations. The GPM speedups over the unaccelerated VNM range between 4.3 and 6.3. Moreover, the speedup ratio achieved with the GPM acceleration increases with an increasing dominance ratio of the problem since the required number of outer iterations increases with an increasing dominance ratio.

Two-stage iterations based on composite splittings for rectangular linear systems

In this paper, we introduce a two-stage method to solve rectangular linear systems that exhibits faster convergence than typical stationary iterative methods. Under suitable conditions, we prove convergence of the new method. The number of outer iterations can be reduced by using a few significant number of inner iterations for efficient computations. Further, we perform a comparison analysis, and establish that a higher number of inner iterations ensures a smaller spectral radius of the global iteration matrix. We also discuss the uniqueness of a proper splitting, and illustrate different comparison theorems for different subclasses of proper splittings.

Prediction of gas–liquid flow in an annular gap bubble column using a bi-dispersed Eulerian model

We present and discuss numerical results from simulations of the air–water flow in an annular gap bubble column of 0.24m internal diameter, at air superficial velocities ranging from 0.004m/s to 0.225m/s, covering the homogeneous and heterogeneous flow regimes. A bi-dispersed Eulerian model is implemented to account for both the stabilizing and destabilizing effects of small and large bubbles. Sensitivity studies on the mesh element size, time step size and number of outer iterations per time step are performed and most appropriate simulation parameters and mesh are used to predict the gas holdup curve. Comparison with two mono-dispersed models is provided to emphasize the necessity of a bi-dispersed approach for the accurate prediction of the homogeneous flow regime, given the poly-dispersed nature of the flow investigated. Two different approaches for the characterization of the small and large bubbles groups are also discussed. We found that the relative amount of small bubbles is an important input parameter for the present model and can be provided using available empirical correlations or experimental data. The results obtained from the simulations also demonstrated the necessity of a population balance model able to capture the bubbles coalescence and breakup phenomena for the correct prediction of the heterogeneous flow regime.

TSIRM: A two-stage iteration with least-squares residual minimization algorithm to solve large sparse linear and nonlinear systems

In this paper, a two-stage iterative algorithm is proposed to improve the convergence of Krylov based iterative methods, typically those of GMRES variants. The principle of the proposed approach is to build an external iteration over the Krylov method, and to frequently store its current residual (at each GMRES restart for instance). After a given number of outer iterations, a least-squares minimization step is applied on the matrix composed by the saved residuals, in order to compute a better solution and to make new iterations if required. It is proven that the proposal has the same convergence properties as the inner embedded method itself. Several experiments have been performed using the PETSc toolkit (using default parameters in the absence of detail) to solve linear and nonlinear problems. They show good speedups compared to GMRES with up to 16,394 cores with different preconditioners.

Group-Based Alternating Direction Method of Multipliers for Distributed Linear Classification.

The alternating direction method of multipliers (ADMM) algorithm has been widely employed for distributed machine learning tasks. However, it suffers from several limitations, e.g., a relative low convergence speed, and an expensive time cost. To this end, in this paper, a novel method, namely the group-based ADMM (GADMM), is proposed for distributed linear classification. In particular, to accelerate the convergence speed and improve global consensus, a group layer is first utilized in GADMM to divide all the slave nodes into several groups. Then, all the local variables (from the slave nodes) are gathered in the group layer to generate different group variables. Finally, by using a weighted average method, the group variables are coordinated to update the global variable (from the master node) until the solution of the global problem is reached. According to the theoretical analysis, we found that: 1) GADMM can mathematically converge at the rate , where is the number of outer iterations and 2) by using the grouping methods, GADMM can improve the convergence speed compared with the distributed ADMM framework without grouping methods. Moreover, we systematically evaluate GADMM on four publicly available LIBSVM datasets. Compared with disADMM and stochastic dual coordinate ascent with alternating direction method of multipliers-ADMM, for distributed classification, GADMM is able to reduce the number of outer iterations, which leads to faster convergence speed and better global consensus. In particular, the statistical significance test has been experimentally conducted and the results validate that GADMM can significantly save up to 30% of the total time cost (with less than 0.6% accuracy loss) compared with disADMM on large-scale datasets, e.g., webspam and epsilon.

Parallel Jacobi–Davidson with block FSAI preconditioning and controlled inner iterations

SummaryThe Jacobi–Davidson (JD) algorithm is considered one of the most efficient eigensolvers currently available for non‐Hermitian problems. It can be viewed as a coupled inner‐outer iteration, where the inner one expands the search subspace and the outer one reduces the eigenpair residual. One of the difficulties in the JD efficient use stems from the definition of the most appropriate inner tolerance, so as to avoid useless extra work and keep the number of outer iterations under control. To this aim, the use of an efficient preconditioner for the inner iterative solver is of paramount importance. The present paper describes a fresh implementation of the JD algorithm with controlled inner iterations and block factorized sparse approximate inverse preconditioning for non‐Hermitian eigenproblems in a parallel computational environment. The algorithm performance is investigated by comparison with a freely available software package such as SLEPc. The results show that combining the inner tolerance control with an efficient preconditioning technique can allow for a significant improvement of the JD performance, preserving a good scalability. Copyright © 2016 John Wiley & Sons, Ltd.

A Symmetric Reduction of the NT Direction

A stable symmetrization of the linear systems arising in interior-point methods for solving linear programs is introduced. A comparison of the condition numbers of the resulting interior-point linear systems with other commonly used approaches indicates that the new approach may be best suitable for an iterative solution. It is shown that there is a natural generalization of this symmetrization to the Nesterov--Todd (NT) search direction for solving semidefinite programs. The generalization includes a novel pivoting strategy to minimize the norm of the right and side and heavily relies on the symmetry properties of the NT direction. The search directions generated by iterative solvers typically have fairly low relative accuracy. Nevertheless, in some preliminary numerical examples, a suitably adapted interior-point approach results in a rather small number of outer iterations.

On R-linear convergence of semi-monotonic inexact augmented Lagrangians for bound and equality constrained quadratic programming problems with application

New convergence results for a variant of the inexact augmented Lagrangian algorithm SMALBE [Z. Dostál, An optimal algorithm for bound and equality constrained quadratic programming problems with bounded spectrum, Computing 78 (2006) 311–328] for the solution of strictly convex bound and equality constrained quadratic programming problems are presented. The algorithm SMALBE-M presented here uses a fixed regularization parameter and controls the precision of the solution of auxiliary bound constrained problems by a multiple of the norm of violation of the equality constraints and a constant which is updated in order to enforce the increase of Lagrangian function. A nice feature of SMALBE-M is its capability to find an approximate solution of important classes of problems in a number of iterations that is independent of the conditioning of the equality constraints. Here we prove the R-linear rate of convergence of the outer loop of SMALBE-M for any positive regularization parameter after the strong active constraints of the solution are identified. The theoretical results are illustrated by solving two benchmarks, including the contact problem of elasticity discretized by two million of nodal variables. The numerical experiments indicate that the inexact solution of auxiliary problems in the inner loop results in a very small increase of the number of outer iterations as compared with the exact algorithm. The results do not assume independent equality constraints and remain valid when the solution is dual degenerate.

Stable discretization of poroelasticity problems and efficient preconditioners for arising saddle point type matrices

Poroelastic models arise in reservoir modeling and many other important applications. Under certain assumptions, they involve a time-dependent coupled system consisting of Navier–Lame equations for the displacements, Darcy’s flow equation for the fluid velocity and a divergence constraint equation. Stability for infinite time of the continuous problem and, second and third order accurate, time discretized equations are shown. Methods to handle the lack of regularity at initial times are discussed and illustrated numerically. After discretization, at each time step this leads to a block matrix system in saddle point form. Mixed space discretization methods and a regularization method to stabilize the system and avoid locking in the pressure variable are presented. A certain block matrix preconditioner is shown to cluster the eigenvalues of the preconditioned matrix about the unit value but needs inner iterations for certain matrix blocks. The strong clustering leads to very few outer iterations. Various approaches to construct preconditioners are presented and compared. The sensitivity of the number of outer iterations to the stopping accuracy of the inner iterations is illustrated numerically.

Robust Multiscale Iterative Solvers for Nonlinear Flows in Highly Heterogeneous Media

In this paper, we study robust iterative solvers for finite element systems re- sulting in approximation of steady-state Richards' equation in porous media with highly heterogeneous conductivity fields. It is known that in such cases the contrast, ratio between the highest and lowest values of the conductivity, can adversely affect the per- formance of the preconditioners and, consequently, a design of robust preconditioners is important for many practical applications. The proposed iterative solvers consist of two kinds of iterations, outer and inner iterations. Outer iterations are designed to handle nonlinearities by linearizing the equation around the previous solution state. As a result of the linearization, a large-scale linear system needs to be solved. This linear system is solved iteratively (called inner iterations), and since it can have large variations in the coefficients, a robust preconditioner is needed. First, we show that under some assump- tions the number of outer iterations is independent of the contrast. Second, based on the recently developed iterative methods, we construct a class of preconditioners that yields convergence rate that is independent of the contrast. Thus, the proposed itera- tive solvers are optimal with respect to the large variation in the physical parameters. Since the same preconditioner can be reused in every outer iteration, this provides an additional computational savings in the overall solution process. Numerical tests are presented to confirm the theoretical results.

Verification of variable-density flow solvers using manufactured solutions

The method of manufactured solutions (MMS) is used to verify the convergence properties of a low-Mach number, variable-density flow code. Three MMS problems relevant to combustion applications are presented and tested on a variety of structured and unstructured grids. Several issues are investigated, including the use of tabulated state properties (i.e., density) and the effect of sub-iterations in the time-advancement method. The MMS implementations provide a quantitative framework to evaluate the impact of these practices on the code’s convergence and order-of-accuracy. Simulation results show that linear interpolation of the equation-of-state causes numerical fluctuations that impede convergence and reduce accuracy. Likewise, the sub-iterative time-advancement scheme requires a significant number of outer iterations to subdue splitting errors in highly nonlinear combustion problems. These findings highlight the importance of careful code and solution verification in the simulation of variable-density flows.

Development of a Stokes flow solver robust to large viscosity jumps using a Schur complement approach with mixed precision arithmetic

We develop an iterative solution technique for solving Stokes flow problems with smooth and discontinuous viscosity structures using a three dimensional, staggered grid finite difference discretization. Two preconditioned iterative methodologies are applied to the saddle point arising from the discrete Stokes problem. They consist of a velocity–pressure coupled approach (FC) and a decoupled, Schur complement approach (SC). Within both of these methods, we utilize either the scaled BFBt, or an identity matrix scaled by the local cell viscosity (LV) to define a preconditioner for the Schur complement. Additionally, we propose to use a mixed precision Krylov kernel to improve the convergence by reducing round-off error. In this approach, standard double precision is used during the application of the preconditioner, whilst higher precision arithmetic is used to define the matrix vector product, dot products and norms required by the Krylov method. In our Krylov kernel, we utilize quad precision arithmetic which is emulated via the double–double precision method. We consider several simplified geodynamic problems with a viscosity contrast to demonstrate the robustness and scalability of our solution methods. Through a careful choice of stopping conditions, we are able to quantitatively compare the residuals between the SC and FC approaches. We examine the trade-off relationship between the number of outer iterations required for convergence, and the computational cost per iteration, for the each solution methods. We find that it is advantageous to use the FC approach utilizing relaxed tolerances for solution of the sub-problems, combined with the LV preconditioner. We also observed that in general, the SC approach is more robust than FC and that BFBt is more robust than LV when used in our numerical experimental. In addition, our mixed precision method produces improved convergence rates of Arnoldi type Krylov subspace methods without a drastic increasing the computational time. The usage of a high precision Krylov kernel is found to be useful for the solver associated with the velocity sub-problem.

Development and verification of an MOC code employing assembly modular ray tracing and efficient acceleration techniques

In this paper, we report the development and verification of a method of characteristics (MOC) code, PEACH, at Shanghai Jiao Tong University. Both the usual flat-source step characteristics (SC) scheme and the linear source (LS) approximation scheme are adopted for the tracking calculation along the neutron trajectory. The assembly-based modular ray tracing (AMRT) technique that possesses a good geometric flexibility and high efficiency is employed, which makes PEACH able to deal with practical LWR assembly and core problems. Moreover, in order to reduce the computational time of the MOC iteration process, both the multi/few-group two-level cell-based coarse mesh finite difference (CMFD) acceleration and the exponential function interpolation technique are used. This results in a significant acceleration. Numerical results for the OECD NEA C5G7 MOX benchmark problem and a 69-group BWR mini-core problem demonstrate that PEACH is accurate and efficient. Numerical results also demonstrate that the LS scheme is more efficient than the SC scheme, taking less time and system memory to generate results of comparable accuracy. In addition, we find that MOC with CMFD acceleration always converges with almost the same number of outer iterations regardless of the physical problem size and the discretization parameters used. This shows an ideal linear relationship between the run time and the geometric size of the problem.

Integration of local–global upscaling and grid adaptivity for simulation of subsurface flow in heterogeneous formations

We propose a methodology, called multilevel local–global (MLLG) upscaling, for generating accurate upscaled models of permeabilities or transmissibilities for flow simulation on adapted grids in heterogeneous subsurface formations. The method generates an initial adapted grid based on the given fine-scale reservoir heterogeneity and potential flow paths. It then applies local–global (LG) upscaling for permeability or transmissibility [7], along with adaptivity, in an iterative manner. In each iteration of MLLG, the grid can be adapted where needed to reduce flow solver and upscaling errors. The adaptivity is controlled with a flow-based indicator. The iterative process is continued until consistency between the global solve on the adapted grid and the local solves is obtained. While each application of LG upscaling is also an iterative process, this inner iteration generally takes only one or two iterations to converge. Furthermore, the number of outer iterations is bounded above, and hence, the computational costs of this approach are low. We design a new flow-based weighting of transmissibility values in LG upscaling that significantly improves the accuracy of LG and MLLG over traditional local transmissibility calculations. For highly heterogeneous (e.g., channelized) systems, the integration of grid adaptivity and LG upscaling is shown to consistently provide more accurate coarse-scale models for global flow, relative to reference fine-scale results, than do existing upscaling techniques applied to uniform grids of similar densities. Another attractive property of the integration of upscaling and adaptivity is that process dependency is strongly reduced, that is, the approach computes accurate global flow results also for flows driven by boundary conditions different from the generic boundary conditions used to compute the upscaled parameters. The method is demonstrated on Cartesian cell-based anisotropic refinement (CCAR) grids, but it can be applied to other adaptation strategies for structured grids and extended to unstructured grids.