文件列表:

generateMGTestSequence.m (1021, 2012-02-24)
generateNet.asv (970, 2012-02-24)
generateNet.m (904, 2012-02-24)
learn.asv (8521, 2012-02-24)
learn.m (8322, 2012-02-24)
Lyapunov.asv (4369, 2012-02-24)
NRMSE84List.mat (488, 2012-02-24)
test.asv (3717, 2012-02-24)
test.m (3716, 2012-02-24)
averageModelTest.asv (661, 2012-02-24)
averageModelTest.m (726, 2012-02-24)
bestModel.mat (147880, 2012-02-24)
continueTrajectory.asv (4196, 2012-02-24)
continueTrajectory.m (4196, 2012-02-24)
generateMGSequence.asv (1942, 2012-02-24)
generateMGSequence.m (1942, 2012-02-24)

Created by Herbert Jaeger, Dec. 4, 2003 This directory contains Matlab files used for training and testing 1000-unit ESN networks on the MG attractor (tau = 17), essentially as described in the Science submission, but with improved parameter settings, so the results are significantly better than reported in the submission. The script averageModelTest.m governs everything. If you call it, first set the numberOftests variable at the top of that script to the number of tests (one test = learning an ESN on randomly generated training data, testing it on numberOfTrials many randomly generated test sequences, where numberOfTrials is a variable that can be set inside the script test.m . Default is numberOfTrials = 50. Each trial is an iterated prediction for 84 MG time units. A run of averageModelTest.m creates and saves a list NRMSE84List of length numberOfTrials, containing the normalized root mean square errors of the tests. A caveat: the random generation of sparse network weight matrices sometimes creates a singular matrix. This is not catched. It happens rarely, but if it happens, the script averageModelTest.m aborts. Besides this, graphical output for prediction runs can be obtained from calling continueTrajectory.m . Note that you first have to execute one call to learn.m to create a learnt model. If you want to call learn.m outside the script averageModelTest.m, uncomment the first line in learn.m. The decisive parameters (network size, noiselevel during training, spectral radius, input scaling and shifting) have been roughly optimized by hand, but further optimization would very likely still significantly improve the average model quality.

近期下载者：

相关文件：

评论：[我要评论] [举报此文件]

收藏者：