Riedmiller, M., RPROP -- Description and Implementation Details, University of Karlsruhe (1994).  Home/Search   Document Details and Download   Summary   Related Articles   Check

....with 50 neurons, another hidden layer with 25 neurons, and 1 output neuron was selected as LM, and used to learn a 2D sinc z # ### # =### #. Initialization of this neural network was performed using a scheme recommended in [7] and the network was trained with the RPROP algorithm [9, 10], which is one of the algorithms that is currently implemented in the Matlab Neural Networks Package. Since our interest in this work is to demonstrate the ability of the present incremental learning algorithm in testing the generalization level, the specific training procedure selected for the ....

Riedmiller, M., RPROP -- Description and Implementation Details, University of Karlsruhe (1994).

....to step by step minimize the di erence between the actual output vector of the network and the desired output vector, such as BP in its di erent versions (like, for instance QuickProp by Fahlman et Preprint submitted to Elsevier Preprint 25 April 2000 al. 1] and RPROP by Riedmiller and Braun [2,3]) is widely used as training mechanism, as well as other evolutionary approaches [4 8] However, even as these BP methods are successfully used in many elds, especially for pattern recognition, due to its learning ability, it does encounter certain diculties in practice: 1) convergence tends to ....

....one gene. The algorithm was run for a xed number of generations. When evaluating each individual of the population to obtain its tness a limit of epochs was established. We have used the BP variant known as the perceptron training algorithm QuickProp [1] This algorithm, together with RPROP [2,3] BP variant, is one of the best avoiding local minima (one of the problems of BP) because the magnitude of the change in the weights (the step size) is not a function of the magnitude of the gradient. The QuickProp algorithm, together with GA, improves the probability of avoiding local minima. 3 ....

M. Riedmiller. RPROP: Description and Implementation Details. Technical report, University of Karlsruhe, 1994.

....MLPs and CasCor. For a fair comparison to the dynamic networks, an appropriate MLP network, with its architecture and parameters hand tuned [8] had to be found for each benchmark. Note that the number of runs needed to obtain these optimized MLPs is not reported here. Resilient Propagation (RP) [6] and Quickprop (QP) 5] were used as learning paradigms. The neural network simulator used was FAST [7] which contained the MLP and CasCor algorithms. 10 runs with both Rprop and Quickprop were made for each of the MLPs, CasCor, and the FlexNet flavors (named after their connection strategy) ....

Riedmiller, M.: "Rprop - Description and Implementation Details", Technical Report, Institut fr Logik, Komplexitt und Deduktionssyteme, Universitt Karlsruhe, 1994.

....Hwang et al. explain these results by pointing out that the use of the correlation measure in Cascor forces the hidden units to saturate, which produces jagged edges in the network outputs. THE CASPER ALGORITHM Casper uses a modified version of the RPROP algorithm (Riedmiller and Braun, 1993; Riedmiller, 1994) for network training. RPROP is a gradient descent algorithm which uses separate adaptive learning rates for each weight. Each weight begins with an initial learning rate, which is then adapted depending on the sign of the error gradient seen by the weight as it traverses the error surface. The ....

Riedmiller, M. (1994) Rprop - Description and Implementation Details, Technical Report, University of Karlsruhe.

....or if in locally quadratic regions second order information is neglected, only poor improvements can be expected for this step. In this paper we propose learning algorithms which dynamically include second order information. These algorithms, which we call hybrid algorithms are based on Rprop [RB94], a Manhattan Learning algorithm. In minimum regions, were the error surface is assumed to be locally quadratic, second order information is introduced by using either one dimensional secant steps (as in Quickprop [Fah88] or by evaluating the diagonal terms of the hessian matrix. The hybrid ....

....on Adaptive Solutions CNAPS, a SIMD computer with 256 processors. On this machine we benchmarked the algorithms against four of the Carnegie Mellon real wold benchmarks. 2 The RPROP Learning Algorithm The Rprop (resilient backpropagation) algorithm, developed by M. Riedmiller and H. Braun [RB94], is an adaptive step algorithm using independent learning rates for each weight. The weights are updated by just taking care of the sign of the gradients components rE(w lij ) not of its magnitude (Manhattan Learning) which very well handles the problem of flat spots or very steep descents of ....

M. Riedmiller and H. Braun. RPROP -- Description and Implementation Details. Technical Report. Universitat Karlsruhe, 1994.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC