We want to evolve a neural network that reacts on objects in its environment.

Eg., a layered ANN, with one layer of inputs (sensors), a hidden layer and an
output layer doing some control.

A simple example would be an ANN-controlled system trying to detect other objects
in a 2D world (like an virtual fish looking for the nearest food source in a 2D
aquarium).
The x and y coordinates of the objects are given and describe the objects'
positions (including properties like distance and angle) perfectly, but should we
feed this information just into the ANN via several "x" and "y" neuron?
Or would it be better to feed the information as "angle" and "radius"? Both ways
would most likely require a lot of neuronal complexity to decode the actual
information in the ANN.

Or, a different idea would be to create a set of n neurons (e.g. that
decomposes the 360 degree view range into n sectors - whenever an object is in a
sector, the particular neuron would be fed with the distance to it. The last
method seems to be more promising, but we are loosing exactness even though we
need a lot more neurons. So we are not sure if this is the right approach..

Has anyone an idea, how this interfacing should be done? We are open for
comments, critics and literature suggestions

On Oct 5, 10:47 am, Peter Bencsik <pbenc... at (no spam) web.de> wrote:
We want to evolve a neural network that reacts on objects in its environment.

Eg., a layered ANN, with one layer of inputs (sensors), a hidden layer and an
output layer doing some control.

A simple example would be an ANN-controlled system trying to detect other objects
in a 2D world (like an virtual fish looking for the nearest food source in a 2D
aquarium).
The x and y coordinates of the objects are given and describe the objects'
positions (including properties like distance and angle) perfectly, but should we
feed this information just into the ANN via several "x" and "y" neuron?
Or would it be better to feed the information as "angle" and "radius"? Both ways
would most likely require a lot of neuronal complexity to decode the actual
information in the ANN.

Or, a different idea would be to create a set of n neurons (e.g. that
decomposes the 360 degree view range into n sectors - whenever an object is in a
sector, the particular neuron would be fed with the distance to it. The last
method seems to be more promising, but we are loosing exactness even though we
need a lot more neurons. So we are not sure if this is the right approach.

Has anyone an idea, how this interfacing should be done? We are open for
comments, critics and literature suggestions :-)


The most realistic model would react to noisy estimates of bearing,
range
and their rates of change.

Hope this helps.

Greg

Greg schrieb:

On Oct 5, 10:47 am, Peter Bencsik <pbenc... at (no spam) web.de> wrote:
We want to evolve a neural network that reacts on objects in its environment.

Eg., a layered ANN, with one layer of inputs (sensors), a hidden layer and an
output layer doing some control.

A simple example would be an ANN-controlled system trying to detect other objects
in a 2D world (like an virtual fish looking for the nearest food source in a 2D
aquarium).
The x and y coordinates of the objects are given and describe the objects'
positions (including properties like distance and angle) perfectly, but should we
feed this information just into the ANN via several "x" and "y" neuron?
Or would it be better to feed the information as "angle" and "radius"? Both ways
would most likely require a lot of neuronal complexity to decode the actual
information in the ANN.

Or, a different idea would be to create a set of n neurons (e.g. that
decomposes the 360 degree view range into n sectors - whenever an object is in a
sector, the particular neuron would be fed with the distance to it. The last
method seems to be more promising, but we are loosing exactness even though we
need a lot more neurons. So we are not sure if this is the right approach.

Has anyone an idea, how this interfacing should be done? We are open for
comments, critics and literature suggestions :-)

The most realistic model would react to noisy estimates of bearing,
range
and their rates of change.

In the model we will later introduce noise.
But our question is, should we feed bearing/range or preprocess the data manually

and feed it in a different format which might be easier to process for the ANN

On Oct 5, 11:30 am, Peter Bencsik <pbenc... at (no spam) web.de> wrote:

But our question is, should we feed bearing/range or preprocess the data manually

MANUALLY??.. .Ball point pens or pencils with erasers?

..or do you reall mean "automatically" (e,g, APPLE I or COMMODORE).

and feed it in a different format which might be easier to process for the ANN

In radar/sonar tracking noisy RAEVr inputs are typically converted to
noisy
XYZVxVyVz for more accurate processing. The results are converted
back
to RAE for the benefit of human understanding.

Greg schrieb:

On Oct 5, 11:30 am, Peter Bencsik <pbenc... at (no spam) web.de> wrote:

[...]

But our question is, should we feed bearing/range or preprocess the data manually

MANUALLY??.. .Ball point pens or pencils with erasers?

..or do you reall mean "automatically"  (e,g, APPLE I or COMMODORE).

sorry, inexact writing. We meant with a (handcrafted) preprocessing program.



and feed it in a different format which might be easier to process for the ANN

In radar/sonar tracking noisy RAEVr inputs are typically converted to
noisy
XYZVxVyVz for more accurate processing. The results are converted
back
to RAE for the benefit of human understanding.

what does RAEVr inputs mean?

this radar/sonar tracking you refer to, has it been implemented with a neural
network? So the inputs to the neural network have been X,Y,Z,Vx,Vy,Vz, right?
What was the output?

If it was done with ANN, could you provide a reference to this application?