R12 Release Notes

Chapter 32
Neural Network Toolbox 4.0 Release Notes

New Features

This section introduces the new features and enhancements added in the Neural Network Toolbox 4.0 since the Neural Network Toolbox 3.0.1 (Release 11.0).

Control System Applications

A new Control Systems chapter of the Neural Network Toolbox User's Guide presents three practical control systems applications:

• Network model predictive control
• Model reference adaptive control
• Feedback linearization controller

Visual Interface

A visual interface has been added to the toolbox. This interface allows you to:

• Create networks
• Enter data into the visual interface
• Initialize, train, and simulate networks
• Export the training results from the visual interface to the command line workspace
• Import data from the command line workspace to the visual interface

To open the Network/Data Manager window, type nntool.

Note    See Platform Limitations for HP and IBM for information about platform-specific limitations for features described in this section.

New Training Functions

The Neural Network Toolbox now has four training algorithms that apply weight and bias learning rules. One algorithm applies the learning rules in batch mode. Three algorithms apply learning rules in three different incremental modes:

• trainb - Batch training function
• trainc - Cyclical order incremental training function
• trainr - Random order incremental training function
• trains - Sequential order incremental training function

All four functions present the whole training set in each epoch (pass through the entire input set).

Note    We no longer recommend using trainwb and trainwb1, which have been replaced by trainb and trainr. The function trainr differs from trainwb1 in that trainwb1 only presented a single vector each epoch instead of going through all vectors, as is done by trainr.

These new training functions are relatively fast because they generate M-code. The functions trainb, trainc, trainr, and trains all generate a temporary M-file consisting of specialized code for training the current network in question.

Design of General Linear Networks

The function newlind now allows you to design linear networks with multiple inputs, outputs, and input delays.

Improved Early Stopping

Early stopping can now be used in combination with Bayesian regularization. In some cases this can improve the generalization capability of the trained network.

Generalization and Speed Benchmarks

Generalization benchmarks comparing the performance of Bayesian regularization and early stopping are provided. Also included are speed benchmarks, which compare the speed of convergence of the various training algorithms on a variety of problems in pattern recognition and function approximation. These benchmarks can aid you in selecting the appropriate algorithm for your problem.

Demonstration of a Sample Training Session

A new demonstration that illustrates a sample training session is included in the "Backpropagation" chapter of the Neural Network Toolbox User's Guide. A sample training session script is also provided. You can modify this script to fit your problem.

Known Software and Documentation Problems

Platform Limitations for HP and IBM