readme.txt

演示在角色扮演游戏中如何利用人工神经网络进行智能分类的训练

=======================================================================
gpdev.net - Backprop Neural Network Sample Application
=======================================================================

Overview:
=========

The sample application demonstrates how a neural-network can be used
in a Role Playing Game.

A Non Player Character (NPC) is trained to classify the player's character into
one of three classes: Fighter, Wizard or Thief.

The application contains 2 tabs - Train and Test.
From the Train tab you can control NPC parameters (Intelligence and Experience),
set the neural-net parameters (learning rate and momentum), view the set of training
patterns and train the neural net.


Train tab
=========

The list on the right side of the dialog displays the set of training patterns.
The Clothing and Weapon columns are the inputs to the network, mapped to 
words to make it more intuitive for us humans (non-humans please don't be offended :)

The Fighter, Wizard, Thief columns are the desired outputs and represent the chance
(between 0 - 1) that the  player's character belongs to this class.
For example, if the Wizard column contains 0.9 then most likely the player is a Wizard.

The icon in the Id column changes from an X (bad pattern) to a V (good pattern) during
training to show the status of each pattern.
A pattern is considered good when the difference between the network outputs
and the pattern's desired outputs is beneath a certain threshold

The NPC Intelligence slider controls the error tolerance of the net during training.
High intelligence gives a tolerance of 0.1 meaning the error of each output node needs
to be beneath this threshold for a training pattern to be considered good.

The NPC Experience slider controls the number of training patterns presented to the network.
An experienced NPC has seen a lot in his life, so he has more memories to learn from :)

The Neural Network settings let you set the learning rate and momentum of the network.
Explaining exactly how these values affect training is beyond the scope of this readme file
But you can find a short explanation on my site at http://gpdev.net/NeuroDriver_bpnet.html

Below the training patterns list you can see several parameters during training:
Number of good patterns, number of iterations so far and total network error.



Test tab
========

After you train the network, click on the Test tab to test the network performance.
Set the network inputs with the Clothing and Weapon sliders and click on the Run button
The network outputs will be displayed near each class label (Fighter, Wizard, Thief)
The higher the value the more likely the player belongs to this class.
The "winner" (output with highest value) is depicted in red.

For example, if you set the Clothing slider to "Robe" and the Weapon slider to "Staff"
and then click on the Run button, most likely the winner will be "Wizard".
If you select "Cloak" and "Dagger" the winner will most likely be "Thief".
(ok, so it's a bit stereotyped, so what :)

If you like, you can change the training set, it's in a file called patterns.txt
in the same folder as the executable.
The full source-code is supplied so you can modify the application any way you wish.



Let me know if you have any comments, suggestions or questions...

Gideon Pertzov
http://gpdev.net
[e-mail]





Legal Stuff
-----------

This application is provided as is and comes with NO WARRANTY of any kind. 
While every attempt has been made to provide a stable product 
there can be no guarantee and the author assumes no liability
for any damages resulting from the use or misuse of this software.
You use this software at your own risk and if you do not agree to these 
terms you should delete this application immediately.

Copyright (c) 2003 Gideon Pertzov