A combined direction stochastic approximation algorithm

The stochastic approximation problem is to find some root or minimum of a nonlinear function in the presence of noisy measurements. The classical algorithm for stochastic approximation problem is the Robbins-Monro (RM) algorithm, which uses the noisy negative gradient direction as the iterative direction. In order to accelerate the classical RM algorithm, this paper gives a new combined direction stochastic approximation algorithm which employs a weighted combination of the current noisy negative gradient and some former noisy negative gradient as iterative direction. Both the almost sure convergence and the asymptotic rate of convergence of the new algorithm are established. Numerical experiments show that the new algorithm outperforms the classical RM algorithm.     

A Stochastic Approximation Frame Algorithm with Adaptive Directions

Stochastic approximation problem is to find some root or extremum of a non- linear function for which only noisy measurements of the function are available.The classical algorithm for stochastic approximation problem is the Robbins-Monro (RM) algorithm,which uses the noisy evaluation of the negative gradient direction as the iterative direction.In order to accelerate the RM algorithm,this paper gives a flame algorithm using adaptive iterative directions.At each iteration,the new algorithm goes towards either the noisy evaluation of the negative gradient direction or some other directions under some switch criterions.Two feasible choices of the criterions are pro- posed and two corresponding flame algorithms are formed.Different choices of the directions under the same given switch criterion in the flame can also form different algorithms.We also proposed the simultanous perturbation difference forms for the two flame algorithms.The almost surely convergence of the new algorithms are all established.The numerical experiments show that the new algorithms are promising.     

Adaptive stochastic approximation algorithm

In this paper, stochastic approximation (SA) algorithm with a new adaptive step size scheme is proposed. New adaptive step size scheme uses a fixed number of previous noisy function values to adjust steps at every iteration. The algorithm is formulated for a general descent direction and almost sure convergence is established. The case when negative gradient is chosen as a search direction is also considered. The algorithm is tested on a set of standard test problems. Numerical results show good performance and verify efficiency of the algorithm compared to some of existing algorithms with adaptive step sizes.     

球面上$\ell_1$正则优化的随机临近梯度方法

本文研究球面上的$\ell_1$正则优化问题,其目标函数由一般光滑函数项和非光滑$\ell_1$正则项构成,且假设光滑函数的随机梯度可由随机一阶oracle估计.这类优化问题被广泛应用在机器学习,图像、信号处理和统计等领域.根据流形临近梯度法和随机梯度估计技术,提出一种球面随机临近梯度算法.基于非光滑函数的全局隐函数定理,分析了子问题解关于参数的Lipschtiz连续性,进而证明了算法的全局收敛性.在基于随机数据集和实际数据集的球面$\ell_1$正则二次规划问题、有限和SPCA问题和球面$\ell_1$正则逻辑回归问题上数值实验结果显示所提出的算法与流形临近梯度法、黎曼随机临近梯度法相比CPU时间上具有一定的优越性.     

求解阵列天线自适应滤波问题的一种调比随机逼近算法

提出了求解阵列天线自适应滤波问题的一种调比随机逼近算法.每一步迭代中,算法选取调比的带噪负梯度方向作为新的迭代方向.相比已有的其他随机逼近算法,这个算法不需要调整稳定性常数,在一定程度上解决了稳定性常数选取难的问题.数值仿真实验表明,算法优于已有的滤波算法,且比经典Robbins-Monro （RM）算法具有更好的稳定性.     

A stochastic steepest-descent algorithm

A stochastic steepest-descent algorithm for function minimization under noisy observations is presented. Function evaluation is done by performing a number of random experiments on a suitable probability space. The number of experiments performed at a point generated by the algorithm reflects a balance between the conflicting requirements of accuracy and computational complexity. The algorithm uses an adaptive precision scheme to determine the number of random experiments at a point; this number tends to increase whenever a stationary point is approached and to decrease otherwise. Two rules are used to determine the number of random experiments at a point; one, in the inner loop of the algorithm, uses the magnitude of the observed gradient of the function to be minimized; and the other, in the outer-loop, uses a measure of accumulated errors in function evaluations at past points generated by the algorithm. Once a stochastic approximation of the function to be minimized is obtained at a point, the algorithm proceeds to generate the next point by using the steepest-descent deterministic methods of Armijo and Polak (Refs. 3, 4). Convergence of the algorithm to stationary points is demonstrated under suitable assumptions.     

Optimal control of bioprocess systems using hybrid numerical optimization algorithms

In the paper, we consider the bioprocess system optimal control problem. Generally speaking, it is very difficult to solve this problem analytically. To obtain the numerical solution, the problem is transformed into a parameter optimization problem with some variable bounds, which can be efficiently solved using any conventional optimization algorithms, e.g. the improved Broyden–Fletcher–Goldfarb–Shanno algorithm. However, in spite of the improved Broyden–Fletcher–Goldfarb–Shanno algorithm is very efficient for local search, the solution obtained is usually a local extremum for non-convex optimal control problems. In order to escape from the local extremum, we develop a novel stochastic search method. By performing a large amount of numerical experiments, we find that the novel stochastic search method is excellent in exploration, while bad in exploitation. In order to improve the exploitation, we propose a hybrid numerical optimization algorithm to solve the problem based on the novel stochastic search method and the improved Broyden–Fletcher–Goldfarb–Shanno algorithm. Convergence results indicate that any global optimal solution of the approximate problem is also a global optimal solution of the original problem. Finally, two bioprocess system optimal control problems illustrate that the hybrid numerical optimization algorithm proposed by us is low time-consuming and obtains a better cost function value than the existing approaches.     

How does a stochastic optimization/approximation algorithm adapt to a randomly evolving optimum/root with jump Markov sample paths

Stochastic optimization/approximation algorithms are widely used to recursively estimate the optimum of a suitable function or its root under noisy observations when this optimum or root is a constant or evolves randomly according to slowly time-varying continuous sample paths. In comparison, this paper analyzes the asymptotic properties of stochastic optimization/approximation algorithms for recursively estimating the optimum or root when it evolves rapidly with nonsmooth (jump-changing) sample paths. The resulting problem falls into the category of regime-switching stochastic approximation algorithms with two-time scales. Motivated by emerging applications in wireless communications, and system identification, we analyze asymptotic behavior of such algorithms. Our analysis assumes that the noisy observations contain a (nonsmooth) jump process modeled by a discrete-time Markov chain whose transition frequency varies much faster than the adaptation rate of the stochastic optimization algorithm. Using stochastic averaging, we prove convergence of the algorithm. Rate of convergence of the algorithm is obtained via bounds on the estimation errors and diffusion approximations. Remarks on improving the convergence rates through iterate averaging, and limit mean dynamics represented by differential inclusions are also presented. The research of G. Yin was supported in part by the National Science Foundation under DMS-0603287, in part by the National Security Agency under MSPF-068-029, and in part by the National Natural Science Foundation of China under #60574069. The research of C. Ion was supported in part by the Wayne State University Rumble Fellowship. The research of V. Krishnamurthy was supported in part by NSERC (Canada).     

A modified Hooke and Jeeves algorithm with likelihood ratio performance extrapolation for simulation optimization

The Hooke and Jeeves algorithm (HJ) is a pattern search procedure widely used to optimize non-linear functions that are not necessarily continuous or differentiable. The algorithm performs repeatedly two types of search routines; an exploratory search and a pattern search. The HJ algorithm requires deterministic evaluation of the function being optimized. In this paper we consider situations where the objective function is stochastic and can be evaluated only through Monte Carlo simulation. To overcome the problem of expensive use of function evaluations for Monte Carlo simulation, a likelihood ratio performance extrapolation (LRPE) technique is used. We extrapolate the performance measure for different values of the decision parameters while simulating a single sample path from the underlying system. Our modified Hooke and Jeeves algorithm uses a likelihood ratio performance extrapolation for simulation optimization. Computational results are provided to demonstrate the performance of the proposed modified HJ algorithm.     

Gradient Estimation Schemes for Noisy Functions

In this paper, we analyze different schemes for obtaining gradient estimates when the underlying functions are noisy. Good gradient estimation is important e.g. for nonlinear programming solvers. As error criterion, we take the norm of the difference between the real and estimated gradients. The total error can be split into a deterministic error and a stochastic error. For three finite-difference schemes and two design of experiments (DoE) schemes, we analyze both the deterministic errors and stochastic errors. We derive also optimal stepsizes for each scheme, such that the total error is minimized. Some of the schemes have the nice property that this stepsize minimizes also the variance of the error. Based on these results, we show that, to obtain good gradient estimates for noisy functions, it is worthwhile to use DoE schemes. We recommend to implement such schemes in NLP solvers.We thank our colleague Jack Kleijnen for useful remarks on an earlier version of this paper and Gül Gürkan for providing us with relevant literature. Moreover, we thank the anonymous referee for valuable remarks.     

Descent Direction Stochastic Approximation Algorithm with Adaptive Step Sizes

A stochastic approximation (SA) algorithm with new adaptive step sizes for solving unconstrained minimization problems in noisy environment is proposed. New adaptive step size scheme uses ordered statistics of fixed number of previous noisy function values as a criterion for accepting good and rejecting bad steps. The scheme allows the algorithm to move in bigger steps and avoid steps proportional to $1/k$ when it is expected that larger steps will improve the performance. An algorithm with the new adaptive scheme is defined for a general descent direction. The almost sure convergence is established. The performance of new algorithm is tested on a set of standard test problems and compared with relevant algorithms. Numerical results support theoretical expectations and verify efficiency of the algorithm regardless of chosen search direction and noise level. Numerical results on problems arising in machine learning are also presented. Linear regression problem is considered using real data set. The results suggest that the proposed algorithm shows promise.     

Learning gradients by a gradient descent algorithm

We propose a stochastic gradient descent algorithm for learning the gradient of a regression function from random samples of function values. This is a learning algorithm involving Mercer kernels. By a detailed analysis in reproducing kernel Hilbert spaces, we provide some error bounds to show that the gradient estimated by the algorithm converges to the true gradient, under some natural conditions on the regression function and suitable choices of the step size and regularization parameters.     

Short-term cascaded hydroelectric system scheduling based on chaotic particle swarm optimization using improved logistic map

In order to solve the model of short-term cascaded hydroelectric system scheduling, a novel chaotic particle swarm optimization (CPSO) algorithm using improved logistic map is introduced, which uses the water discharge as the decision variables combined with the death penalty function. According to the principle of maximum power generation, the proposed approach makes use of the ergodicity, symmetry and stochastic property of improved logistic chaotic map for enhancing the performance of particle swarm optimization (PSO) algorithm. The new hybrid method has been examined and tested on two test functions and a practical cascaded hydroelectric system. The experimental results show that the effectiveness and robustness of the proposed CPSO algorithm in comparison with other traditional algorithms.     

Convergence analysis of gradient descent stochastic algorithms

This paper proves convergence of a sample-path based stochastic gradient-descent algorithm for optimizing expected-value performance measures in discrete event systems. The algorithm uses increasing precision at successive iterations, and it moves against the direction of a generalized gradient of the computed sample performance function. Two convergence results are established: one, for the case where the expected-value function is continuously differentiable; and the other, when that function is nondifferentiable but the sample performance functions are convex. The proofs are based on a version of the uniform law of large numbers which is provable for many discrete event systems where infinitesimal perturbation analysis is known to be strongly consistent.     

Minimax efficient finite-difference stochastic gradient estimators using black-box function evaluations

Standard approaches to stochastic gradient estimation, with only noisy black-box function evaluations, use the finite-difference method or its variants. Though natural, it is open to our knowledge whether their statistical accuracy is the best possible. This paper argues so by showing that central finite-difference is a nearly minimax optimal zeroth-order gradient estimator for a suitable class of objective functions and mean squared risk, among both the class of linear estimators and the much larger class of all (nonlinear) estimators.     

Robust optimal control for a batch nonlinear enzyme-catalytic switched time-delayed process with noisy output measurements

In this paper, we consider a nonlinear switched time-delayed (NSTD) system with an unknown time-varying function describing the batch culture. The output measurements are noisy. According to the actual fermentation process, this time-varying function appears in the form of a piecewise-linear function with unknown kinetic parameters and switching times. The quantitative definition of biological robustness is given to overcome the difficulty of accurately measuring intracellular material concentrations. Our main goal is to estimate these unknown quantities by using noisy output measurements and biological robustness. This estimation problem is formulated as a robust optimal control problem (ROCP) governed by the NSTD system subject to continuous state inequality constraints. The ROCP is approximated as a sequence of nonlinear programming subproblems by using some techniques. Due to the highly complex nature of these subproblems, we propose a hybrid parallel algorithm, based on Nelder–Mead method, simulated annealing and the gradients of the constraint functions, for solving these subproblems. The paper concludes with simulation results.     

Support vector machine classification with indefinite kernels

We propose a method for support vector machine classification using indefinite kernels. Instead of directly minimizing or stabilizing a nonconvex loss function, our algorithm simultaneously computes support vectors and a proxy kernel matrix used in forming the loss. This can be interpreted as a penalized kernel learning problem where indefinite kernel matrices are treated as noisy observations of a true Mercer kernel. Our formulation keeps the problem convex and relatively large problems can be solved efficiently using the projected gradient or analytic center cutting plane methods. We compare the performance of our technique with other methods on several standard data sets.     

A policy gradient method for semi-Markov decision processes with application to call admission control

Solving a semi-Markov decision process (SMDP) using value or policy iteration requires precise knowledge of the probabilistic model and suffers from the curse of dimensionality. To overcome these limitations, we present a reinforcement learning approach where one optimizes the SMDP performance criterion with respect to a family of parameterised policies. We propose an online algorithm that simultaneously estimates the gradient of the performance criterion and optimises it using stochastic approximation. We apply our algorithm to call admission control. Our proposed policy gradient SMDP algorithm and its application to admission control is novel.     

Determination of space‐dependent coefficients from temperature measurements using the conjugate gradient method

In this article, we consider coefficient identification problems in heat transfer concerned with the determination of the space‐dependent perfusion coefficient and/or thermal conductivity from interior temperature measurements using the conjugate gradient method (CGM). We establish the direct, sensitivity and adjoint problems and the iterative CGM algorithm which has to be stopped according to the discrepancy principle in order to reconstruct a stable solution for the inverse problem. The Sobolev gradient concept is introduced in the CGM iterative algorithm in order to improve the reconstructions. The numerical results illustrated for both exact and noisy data, in one‐ and two‐dimensions for single or double coefficient identifications show that the CGM is an efficient and stable method of inversion.     

An algorithm for approximating piecewise linear concave functions from sample gradients

An effective algorithm for solving stochastic resource allocation problems is to build piecewise linear, concave approximations of the recourse function based on sample gradient information. Algorithms based on this approach are proving useful in application areas such as the newsvendor problem, physical distribution and fleet management. These algorithms require the adaptive estimation of the approximations of the recourse function that maintain concavity at every iteration. In this paper, we prove convergence for a particular version of an algorithm that produces approximations from stochastic gradient information while maintaining concavity.