G. Magoulas, V. Plagianakos, and M. Vrahatis. Hybrid methods using evolutionary algorithms for on--line training. In Proc. of the IEEE International Joint Conference on Neural Networks (IJCNN'2001.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....is chosen to compare and evaluate the training algorithms, as an example of practical application, using continuous valued training data that contain random noise. On average, BP trained networks with a 16 8 12 architecture perform better than others according to our previous experiments [40] [41]. Thus, 1000 trained MLPs of this architecture were tested for their generalization capability, using test patterns from 20 subimages of size 128 128, randomly selected from each image. To evaluate the average generalization performance of the MLPs the rule was used, i.e. a test pattern is ....

G. D. Magoulas, V. P. Plagianakos, and M. N. Vrahatis, "Hybrid methods using evolutionary algorithms for on-line training," in Proc. IEEE Int. Joint Conf. Neural Networks, Washington, DC, 2001, pp. 2218--2223.

.... time series is given as: v= lim dlnC(r,m) r ,m d ln(r) To estimate the maximum Lyapunov exponent, 30, which is a measure to estimate the strength of chaos, the following relation is used: logln So I (n = to enat = lim , n nat where n is the divergence of the trajectories over the time [7]. In Figure 1, the seismic events (with magnitude greater than 2 degrees of Richter scale) that occurred in Greece the last twenty years are exhibited. To better analyze the seismic data, we have separated them in four sets, according to the geographical region that the earthquake took place. ....

....used. The aim was to train the ANN to perform short term forecasting. 10,000 consecutive 5 dimensional vectors constitute the training set, while the generalization capability of the network was tested using 1,000 test vectors. The network was trained by a recently proposed learning algorithm [7] to perform one step ahead forecast. The magnitude of the earthquake is forecasted one step at the time and the actual rather than the forecasted magnitude is then used for the next prediction in a forecasting horizon. As shown in Figure 1, the prediction results were satisfactory. 5 Conclusions ....

G.D. Magoulas, V.P. Plagianakos, and M.N. Vrahatis, "Hybrid Methods Using Evolutionary Algorithms for On-line Training", In Proc. of the IEEE Int. Joint Conf on Neural Networks (IJCNN'2000.

....in each of the four frames and contains normal and abnormal samples. In Table 2, the generalization capability of the algorithms, on the test set, is exhibited. Better generalization results for this di#cult problem can be achieved using specially designed training algorithms (see for example [7]) Algorithm Generalization BP 78.09 BPM 78.09 ABP 79.10 NMBBP 83.91 NMBPVS 85.12 Table 2: Generalization Results. Finally, we have tried to determine the speedup of the parallel procedure. Several factors can influence the speedup, such as the network load and the CPU load due to ....

G.D. Magoulas, V.P. Plagianakos, and M.N. Vrahatis, Hybrid Methods Using Evolutionary Algorithms for On--line Training, In Proc. of the IEEE International Joint Conference on Neural Networks (IJCNN'2001), Washington D.C., (2001).

No context found.

G. Magoulas, V. Plagianakos, and M. Vrahatis. Hybrid methods using evolutionary algorithms for on--line training. In Proc. of the IEEE International Joint Conference on Neural Networks (IJCNN'2001.

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