From edwin@maui.cs.ucla.edu Wed Apr 13 19:53:48 EDT 1994
Article: 15722 of comp.ai.neural-nets
Xref: glinda.oz.cs.cmu.edu comp.ai.neural-nets:15722
Newsgroups: comp.ai.neural-nets
Path: honeydew.srv.cs.cmu.edu!nntp.club.cc.cmu.edu!news.mic.ucla.edu!ucla-cs!maui.cs.ucla.edu!edwin
From: edwin@maui.cs.ucla.edu (Edwin Tisdale)
Subject: Neural Network Learns to Fly
Nntp-Posting-Host: maui.cs.ucla.edu
Organization: UCLA, Computer Science Department
Message-ID: <1994Apr01.215242.11174@cs.ucla.edu>
Date: Fri, 01 Apr 94 21:52:42 GMT
Lines: 71

Version 2.5.8 of the Gnu C++ compiler compiles the Artificial Neural Controller
without complaint under SunOs and Linux.  Version 2.1 of the "Sorry, not
implemented" Sun C++ compiler will not accept all of the code generated
by a Fortran to C Converter.  I would appreciate it very much if someone
would try to compile it with the latest version of the Sun C++ compiler.
I would be glad to help work around any problems you might encounter.

The Real Time Recurrent Learning Algorithm was used to train a multi-layer,
feed-forward Artificial Neural Network controller to perform the function
of a Flight Test Maneuver Autopilot for a High Performance Fighter Aircraft
flight simulator.  Derivative Arithmetic replaces the normal floating-point
arithmetic in a conventional computer model so that the partial derivatives
of the state variables at the next time step with respect to the state and
control variables at the current time step are computed along with the normal
result.  These partial derivatives are propagated backward through the layers
of the network and used to update the total partial derivatives of the state
and control variables with respect to the biases and connection weights.
The AIAA Aircraft Controls Design Challenge computer model was converted
from FORTRAN 77 to C++ using A Fortran-to-C Converter so that a C++ class
could be used to implement derivative arithmetic.

The C++ Matrix class is used to implement the Artificial Neural Controller
and the Real Time Recurrent Learning Algorithm.  The main program can be
used to train an Artificial Neural Controller for any nonlinear system
if source code for an accurate computer model is available.

The flight simulator is available via anonymous FTP from `ftp.cs.ucla.edu'.
Get the compressed tarfile, `/pub/simulator.tar.Z', uncompress it and extract
the `simulator' directory from `simulator.tar'.  Then go to the `simulator'
directory and type `make'.

The flight simulator was used to train an Artificial Neural Controller to
fly a 2g coordinated turn at Mach 1 and 20000 feet.  If you want to see
the simulator fly this turn, type `cp network.bak network.old' then type
`src/simulate -v'.  The program will display state and control variables
every 127 time steps.  Read the documentation in the `readme' file if
you want to see it LEARN to fly this maneuver.

Enjoy, Bob Tisdale (edwin@cs.ucla.edu)

P.S.  I hope the following notes will be helpful.

unix% ftp ftp.cs.ucla.edu
Name (ftp.cs.ucla.edu:your_login_ID): anonymous
Password: your_login_ID
ftp> cd pub
ftp> binary
ftp> get simulator.tar.Z
ftp> bye
unix% uncompress simulator.tar.Z
unix% tar xvf simulator.tar
unix% cd simulator
unix% make
unix% cp network.bak network.old
unix% src/simulate -v

Note: ftp.cs.ucla.edu is an alias for internet-address 131.179.128.36
      The C++ Matrix class can be obtained via anonymous FTP from
      the same site in file `/pub/Matrix.tar.Z'.  The `libI77.a' and
      `libF77.a' Fortran libraries can be obtained via anonymous FTP
      from `netlib.att.com:/netlib/f2c/'.

      The Gnu C++ compiler, `g++', and library, `libg++', are required.
      It has been tested on Sun SPARC computers running Solaris but
      should compile and run on any UNIX workstation with a Gnu C++
      compiler and library.  Instructions for installing the Gnu C++
      compiler and libraries can be obtained in either plain text or
      texinfo format via anonymous FTP from `rtfm.mit.edu' from the
      `/pub/usenet/news.answers/g++-FAQ' directory.  I will be glad to
      help anyone who wishes to port this code to other C++ compilers
      if they are willing to make it available to others.