Bayesian network over neural netowrks developed in ETL (BAYONNET-Applet)

Bayesian networks on Neural Networks developed in ETL (BAYONNET-Applet)

Example for Robot Navigation

[Introduction]

This network is working for our Jijo2 robot .
Conditional probabilities are represented by neural networks, which are trained by data sampled in the real environment.
St denotes previous location (position No.0 ... No.11).
St+1 denotes next location (position No.0 ... No.11).
At denotes action of the robot (0=go_forward and 1=go_backward).
Dt+1 denotes sensor value. Displayed values are mean (0) and variance (1) of the gaussian.

[Learning]

Learning conditional probabilities is realized by the newer version of this system.
It uses JDBC mechanism, and training samples are provided by commersial database package (Oracle, MS SQLserver and so on).
We also have a plan to extend this system for so-called Data Mining

[How to play]

1. Wait for the loading conditional probability files (for each node).
2. Set current position as evidence of St, and action as evdence of At
3. The prediction of next position St+1 is obtained by clicking "Estimate" button of St+1.
4. The prediction of sensor value Dt+1 is obtained by clicking "Estimate" button of Dt+1.
5. You can edit original network by "New Node", "Link", "Move" and "Cut" .
6. Finally, pop up windows and BNsimulator are closed by "Quit".

If you can use JAVA available browser, you can play this applet directly.

This Applet is listed in the Java-Repository - a collection of documented Java-ressources,

[Program]

source codes (Applet version)
source codes (Applications(BNserver(TCP/IP) + GUII/F) version)
(% The authors will not take any responsibity about results of this program.)

Research Papers

List with Abstract(in English):