Five different artificial neural network strategies have been
tested to solve a simulated undersea obstacle avoidance problem
for an underwater vehicle.
The primary source of input to the neural networks have been
preprocessed sonar data from a sonar having 60 lobes.
The output control signals from the networks have been the forward-,
vertical-, and horizontal angular velocities of the vehicle.
Evolutionary algorithms have successfully been used to train the
neural networks, and in one case supervised learning has been
used as well. A simulator environment, where the networks have
been tested, has been designed using MATLAB, and the outcome
of the simulations were promising. It turned out that
the simple one layer network structures solved the problem better
than the more complicated network structures did.
A prototype to a program that will be used when implementing
the neural networks on a real autonomous underwater vehicle
has been constructed as well, using the C++ programming
language. The program is easy to operate, using a graphical
user interface, and the object oriented design should make it
easy to include further neural network strategies in the future.
For several reasons, the neural network control program
has not yet been tested on the real autonomous vessel,
and these tests are postponed to a later date.