Parallel evolutionary training algorithms for “hardware-friendly” neural networks

In this paper, Parallel Evolutionary Algorithms for integer weightneural network training are presented. To this end, each processoris assigned a subpopulation of potential solutions. Thesubpopulations are independently evolved in parallel andoccasional migration is employed to allow cooperation betweenthem. The proposed algorithms are applied to train neural networksusing threshold activation functions and weight values confined toa narrow band of integers. We constrain the weights and biases inthe range [–3, 3], thus they can be represented by just 3 bits.Such neural networks are better suited for hardware implementationthan the real weight ones. These algorithms have been designedkeeping in mind that the resulting integer weights require lessbits to be stored and the digital arithmetic operations betweenthem are easier to be implemented in hardware. Another advantageof the proposed evolutionary strategies is that they are capableof continuing the training process ``on-chip', if needed. Ourintention is to present results of parallel evolutionaryalgorithms on this difficult task. Based on the application of theproposed class of methods on classical neural network problems,our experience is that these methods are effective and reliable.     

Extraction of Logical Rules from Neural Networks

Three neural-based methods for extraction of logical rules from data are presented. These methods facilitate conversion of graded response neural networks into networks performing logical functions. MLP2LN method tries to convert a standard MLP into a network performing logical operations (LN). C-MLP2LN is a constructive algorithm creating such MLP networks. Logical interpretation is assured by adding constraints to the cost function, forcing the weights to ±1 or 0. Skeletal networks emerge ensuring that a minimal number of logical rules are found. In both methods rules covering many training examples are generated before more specific rules covering exceptions. The third method, FSM2LN, is based on the probability density estimation. Several examples of performance of these methods are presented.     

Classical and superposed learning for quantum weightless neural networks

A supervised learning algorithm for quantum neural networks (QNN) based on a novel quantum neuron node implemented as a very simple quantum circuit is proposed and investigated. In contrast to the QNN published in the literature, the proposed model can perform both quantum learning and simulate the classical models. This is partly due to the neural model used elsewhere which has weights and non-linear activations functions. Here a quantum weightless neural network model is proposed as a quantisation of the classical weightless neural networks (WNN). The theoretical and practical results on WNN can be inherited by these quantum weightless neural networks (qWNN). In the quantum learning algorithm proposed here patterns of the training set are presented concurrently in superposition. This superposition-based learning algorithm (SLA) has computational cost polynomial on the number of patterns in the training set.     

Cooperative-Competitive Algorithms for Evolutionary Networks Classifying Noisy Digital Images

We describe an efficient method of combining the global search of genetic algorithms (GAs) with the local search of gradient descent algorithms. Each technique optimizes a mutually exclusive subset of the network's weight parameters. The GA chromosome fixes the feature detectors and their location, and a gradient descent algorithm starting from random initial values optimizes the remaining weights. Three algorithms having different methods of encoding hidden unit weights in the chromosome are applied to multilayer perceptrons (MLPs) which classify noisy digital images. The fitness function measures the MLP classification accuracy together with the confidence of the networks.     

递归pi-sigma神经网络的带惩罚项的梯度算法分析

传统的梯度算法存在收敛速度过慢的问题,针对这个问题,提出一种将惩罚项加到传统误差函数的梯度算法以训练递归pi-sigma神经网络,算法不仅提高了神经网络的泛化能力,而且克服了因网络初始权值选取过小而导致的收敛速度过慢的问题,相比不带惩罚项的梯度算法提高了收敛速度。从理论上分析了带惩罚项的梯度算法的收敛性,并通过实验验证了算法的有效性。     

基于反传混沌粒子群训练的前馈神经网络研究

为了解决前馈神经网络训练收敛速度慢、易陷入局部极值及对初始权值依赖性强等缺点, 提出了一种基于反传的无限折叠迭代混沌粒子群优化(ICMICPSO)算法训练前馈神经网络(FNNs)参数。该方法在充分利用BP算法的误差反传信息和梯度信息的基础上, 引入了ICMIC混沌粒子群的概念, 将ICMIC粒子群(ICMICPS)作为全局搜索器, 梯度下降信息作为局部搜索器来调整网络的权值和阈值, 使得粒子能够在全局寻优的基础上对整个空间进行搜索。通过仿真实验与多种算法进行对比, 结果表明在训练和泛化能力上ICMICPSO-BPNN方法明显优于其他算法。     

An iterative method for training multilayer networks with thresholdfunctions

Concerns the problem of finding weights for feed-forward networks in which threshold functions replace the more common logistic node output function. The advantage of such weights is that the complexity of the hardware implementation of such networks is greatly reduced. If the task to be learned does not change over time, it may be sufficient to find the correct weights for a threshold function network off-line and to transfer these weights to the hardware implementation. This paper provides a mathematical foundation for training a network with standard logistic function nodes and gradually altering the function to allow a mapping to a threshold unit network. The procedure is analogous to taking the limit of the logistic function as the gain parameter goes to infinity. It is demonstrated that, if the error in a trained network is small, a small change in the gain parameter will cause a small change in the network error. The result is that a network that must be implemented with threshold functions can first be trained using a traditional back propagation network using gradient descent, and further trained with progressively steeper logistic functions. In theory, this process could require many repetitions. In simulations, however, the weights have be successfully mapped to a true threshold network after a modest number of slope changes. It is important to emphasize that this method is only applicable to situations for which off-line learning is appropriate.     

Combined projection and kernel basis functions for classification in evolutionary neural networks

This paper proposes a hybrid neural network model using a possible combination of different transfer projection functions (sigmoidal unit, SU, product unit, PU) and kernel functions (radial basis function, RBF) in the hidden layer of a feed-forward neural network. An evolutionary algorithm is adapted to this model and applied for learning the architecture, weights and node typology. Three different combined basis function models are proposed with all the different pairs that can be obtained with SU, PU and RBF nodes: product–sigmoidal unit (PSU) neural networks, product–radial basis function (PRBF) neural networks, and sigmoidal–radial basis function (SRBF) neural networks; and these are compared to the corresponding pure models: product unit neural network (PUNN), multilayer perceptron (MLP) and the RBF neural network. The proposals are tested using ten benchmark classification problems from well known machine learning problems. Combined functions using projection and kernel functions are found to be better than pure basis functions for the task of classification in several datasets.     

Neural Networks in Non-Euclidean Spaces

Multilayer Perceptrons (MLPs) use scalar products to compute weighted activation of neurons providing decision borders using combinations of soft hyperplanes. The weighted fun-in activation function may be replaced by a distance function between the inputs and the weights, offering a natural generalization of the standard MLP model. Non-Euclidean distance functions may also be introduced by normalization of the input vectors into an extended feature space. Both approaches influence the shapes of decision borders dramatically. An illustrative example showing these changes is provided.     

Evolving Multilayer Perceptrons

This paper proposes a new version of a method (G-Prop, genetic backpropagation) that attempts to solve the problem of finding appropriate initial weights and learning parameters for a single hidden layer Multilayer Perceptron (MLP) by combining an evolutionary algorithm (EA) and backpropagation (BP). The EA selects the MLP initial weights, the learning rate and changes the number of neurons in the hidden layer through the application of specific genetic operators, one of which is BP training. The EA works on the initial weights and structure of the MLP, which is then trained using QuickProp; thus G-Prop combines the advantages of the global search performed by the EA over the MLP parameter space and the local search of the BP algorithm. The application of the G-Prop algorithm to several real-world and benchmark problems shows that MLPs evolved using G-Prop are smaller and achieve a higher level of generalization than other perceptron training algorithms, such as QuickPropagation or RPROP, and other evolutive algorithms. It also shows some improvement over previous versions of the algorithm.     

Financial decision support using neural networks and support vector machines

Abstract: Bankruptcy prediction and credit scoring are the two important problems facing financial decision support. The multilayer perceptron (MLP) network has shown its applicability to these problems and its performance is usually superior to those of other traditional statistical models. Support vector machines (SVMs) are the core machine learning techniques and have been used to compare with MLP as the benchmark. However, the performance of SVMs is not fully understood in the literature because an insufficient number of data sets is considered and different kernel functions are used to train the SVMs. In this paper, four public data sets are used. In particular, three different sizes of training and testing data in each of the four data sets are considered (i.e. 3:7, 1:1 and 7:3) in order to examine and fully understand the performance of SVMs. For SVM model construction, the linear, radial basis function and polynomial kernel functions are used to construct the SVMs. Using MLP as the benchmark, the SVM classifier only performs better in one of the four data sets. On the other hand, the prediction results of the MLP and SVM classifiers are not significantly different for the three different sizes of training and testing data.     

A novel model by evolving partially connected neural network for stock price trend forecasting

This paper proposes a novel model by evolving partially connected neural networks (EPCNNs) to predict the stock price trend using technical indicators as inputs. The proposed architecture has provided some new features different from the features of artificial neural networks: (1) connection between neurons is random; (2) there can be more than one hidden layer; (3) evolutionary algorithm is employed to improve the learning algorithm and training weights. In order to improve the expressive ability of neural networks, EPCNN utilizes random connection between neurons and more hidden layers to learn the knowledge stored within the historic time series data. The genetically evolved weights mitigate the well-known limitations of gradient descent algorithm. In addition, the activation function is defined using sin(x) function instead of sigmoid function. Three experiments were conducted which are explained as follows. In the first experiment, we compared the predicted value of the trained EPCNN model with the actual value to evaluate the prediction accuracy of the model. Second experiment studied the over fitting problem which occurred in neural network training by taking different number of neurons and layers. The third experiment compared the performance of the proposed EPCNN model with other models like BPN, TSK fuzzy system, multiple regression analysis and showed that EPCNN can provide a very accurate prediction of the stock price index for most of the data. Therefore, it is a very promising tool in forecasting of the financial time series data.     

基于样条插值函数的离散过程神经网络训练*

为解决过程神经元网络不能直接输入离散样本的问题,提出基于样条插值函数的离散过程神经网络训练算法。首先,将离散过程样本按采样点分段,在采样区间内分别构造样本和权值的分段样条函数;然后,计算样本函数和权函数的乘积在采样区间上的积分,并将此积分值提交给网络的隐层过程神经元;最后,在输出层计算网络输出。分别采用一次、二次、三次样条函数,设计了三种不同的算法。实验结果表明：一次样条计算效率高,逼近能力差;三次样条计算效率低,但逼近能力好;二次样条在计算效率和逼近能力两方面都比较理想。因此,二次样条函数是离散过程神经网络的较好选择。     

Application of modified MLP input weights’ matrices: an analysis of sectorial investment distribution in the emerging markets

This paper aims to explore the modified multilayer perceptron (MLP) input weights’ (IW) matrices relating them with the weights of the constituent input determinants. Non-traditional MLP topologies were designed, optimized and compared with other neural networks (NN) and multidimensional linear regression methods and statistically tested. The chosen NN topology directly related the MLP IW matrices with the relative contribution of each input variable. The contribution (weights) of each input variable was estimated in a non-linear manner, which is a novel approach in the investment research domain. This approach was applied to an investigation of sectorial investment distribution in emerging investment markets. To our knowledge, there is no experiment in the field that would focus on the NN mechanisms of sectorial indices (SI) weights estimation in such an experimental setting. In summary, we found apparent correlations between multivariate linear and other NN estimates (like Garson’s, Tchaban’s and SNA methods) having some new results not revealed in the previous research.     

采用资格迹的神经网络学习控制算法

强化学习是解决自适应问题的重要方法,被广泛地应用于连续状态下的学习控制,然而存在效率不高和收敛速度较慢的问题.在运用反向传播(back propagation,BP)神经网络基础上,结合资格迹方法提出一种算法,实现了强化学习过程的多步更新.解决了输出层的局部梯度向隐层节点的反向传播问题,从而实现了神经网络隐层权值的快速更新,并提供一个算法描述.提出了一种改进的残差法,在神经网络的训练过程中将各层权值进行线性优化加权,既获得了梯度下降法的学习速度又获得了残差梯度法的收敛性能,将其应用于神经网络隐层的权值更新,改善了值函数的收敛性能.通过一个倒立摆平衡系统仿真实验,对算法进行了验证和分析.结果显示,经过较短时间的学习,本方法能成功地控制倒立摆,显著提高了学习效率.     

A hybrid linear/nonlinear training algorithm for feedforward neuralnetworks

This paper presents a new hybrid optimization strategy for training feedforward neural networks. The algorithm combines gradient-based optimization of nonlinear weights with singular value decomposition (SVD) computation of linear weights in one integrated routine. It is described for the multilayer perceptron (MLP) and radial basis function (RBF) networks and then extended to the local model network (LMN), a new feedforward structure in which a global nonlinear model is constructed from a set of locally valid submodels. Simulation results are presented demonstrating the superiority of the new hybrid training scheme compared to second-order gradient methods. It is particularly effective for the LMN architecture where the linear to nonlinear parameter ratio is large.     

Boundedness and Convergence of Online Gradient Method with Penalty for Linear Output Feedforward Neural Networks

This paper investigates an online gradient method with penalty for training feedforward neural networks with linear output. A usual penalty is considered, which is a term proportional to the norm of the weights. The main contribution of this paper is to theoretically prove the boundedness of the weights in the network training process. This boundedness is then used to prove an almost sure convergence of the algorithm to the zero set of the gradient of the error function.     

Bio-inspired and gradient-based algorithms to train MLPs: The influence of diversity

This paper has three main goals: (i) to employ two classes of algorithms: bio-inspired and gradient-based to train multi-layer perceptron (MLP) neural networks for pattern classification; (ii) to combine the trained neural networks into ensembles of classifiers; and (iii) to investigate the influence of diversity in the classification performance of individual and ensembles of classifiers. The optimization version of an artificial immune network, named opt-aiNet, particle swarm optimization (PSO) and an evolutionary algorithm (EA) are used as bio-inspired methods to train MLP networks. Besides, the standard backpropagation with momentum (BPM), a quasi-Newton method called DFP and a modified scaled-conjugate gradient (SCGM) are the gradient-based algorithms used to train MLP networks in this work. Comparisons among all the training methods are presented in terms of classification accuracy and diversity of the solutions found. The results obtained suggest that most bio-inspired algorithms deteriorate the diversity of solutions during the search, while immune-based methods, like opt-aiNet, and multiple initializations of standard gradient-based algorithms provide diverse solutions that result in good classification accuracy for the ensembles.     

Optimization of the weights and asymmetric activation function family of neural network for time series forecasting

The use of neural network models for time series forecasting has been motivated by experimental results that indicate high capacity for function approximation with good accuracy. Generally, these models use activation functions with fixed parameters. However, it is known that the choice of activation function strongly influences the complexity and neural network performance and that a limited number of activation functions has been used in general. We describe the use of an asymmetric activation functions family with free parameter for neural networks. We prove that the activation functions family defined, satisfies the requirements of the universal approximation theorem We present a methodology for global optimization of the activation functions family with free parameter and the connections between the processing units of the neural network. The main idea is to optimize, simultaneously, the weights and activation function used in a Multilayer Perceptron (MLP), through an approach that combines the advantages of simulated annealing, tabu search and a local learning algorithm. We have chosen two local learning algorithms: the backpropagation with momentum (BPM) and Levenberg–Marquardt (LM). The overall purpose is to improve performance in time series forecasting.     

粒子群优化算法在神经网络控制中的应用

考虑粒子群优化算法在不确定系统的自适应控制中的应用。神经网络在不确定系统的自适应控制中起着重要作用。但传统的梯度下降法训练神经网络时收敛速度慢,容易陷入局部极小,且对网络的初始权值等参数极为敏感。为了克服这些缺点,提出了一种基于粒子群算法优化的RBF神经网络整定PID的控制策略。首先,根据粒子群算法的基本原理提出了优化得到RBF神经网络输出权、节点中心和节点基宽参数的初值的算法。其次,再利用梯度下降法对控制器参数进一步调节。将传统的神经网络控制与基于粒子群优化的神经网络控制进行了对比,结果表明,后者有更好逼近精度。以PID控制器参数整定为例,对一类非线性控制系统进行了仿真。仿真结果表明基于粒子群优化的神经网络控制具有较强的鲁棒性和自适应能力。