# Using Neural Network as autopilot

Say we have a vehicle (a car or a boat or a helicopter or an airplane etc) which reacts to control inputs from a driver/pilots in a particular way
defined by its current state and control input.
Assume that we have a car travelling on flat road. Our control input would be a vector of Ctrl_Input = [Steering Angle, Gear, Accelerator Value, Brake Value] and current state would be a vector of Current_State = [Velocity X, Velocity Y, Position X, Position Y, Acceleration X, Acceleration Y] the next state of the vehicle after a small time increment "dt" would be Next_State = [Velocity X, Velocity Y, Position X, Position Y, Acceleration X, Acceleration Y]
In another word we have a system with following input and output
Inputs: Ctrl_Inputs Current_State Output: (obtained after "dt" time) Next_State
Assume that we have as a set of Ctrl_Inputs, Current_State and Next_State data obtained in an experiment inwhich a human driver driving the vehicle on a flat road.
Given these data sets we should be able to train a neural network so that it relate inputs to outputs. My questions are as follows; 1- Which neural network topology would be good for learning vehicle behaviour (dynamic model of the vehicle) from above mentioned set of data ? 2- Number of neurons and number of layers required? 3- If we know our current state and our desired state (next state) how we can use the neural network to find required Control_Input vector? 4- Is there any tutorial/demo/code of this kind ? 5- Thanks.
<% if( /^image/.test(type) ){ %>
<% } %>
<%-name%>
have a look "reinforcement learning" which maps input-state to action, at a local level such that the global goals is achieved. its used to control robots etc etc ...
t
On Sat, 24 Apr 2004 01:35:42 +1000, Autopilot wrote:

<% if( /^image/.test(type) ){ %>
<% } %>
<%-name%>
Autopilot wrote:

The EvSail and SailChallenge software packages, downloadable gratis from the URL below, are examples of ANNs "trained" to control a virtual sailboat. Most of this software is applicable to your problem. These are written in C, and are fully documented.
Our ANNS are not actually trained. Instead, a good one is evolved by simulated evolution. The topology is an MLP with one hidden layer and about 40 neurons.
Mitchell Timin
--
"Many are stubborn in pursuit of the path they have chosen, few in
pursuit of the goal." - Friedrich Nietzsche
<% if( /^image/.test(type) ){ %>
<% } %>
<%-name%>
Autopilot wrote:

I trained an F15 simulator to fly coordinated turns at 2g.
<% if( /^image/.test(type) ){ %>
<% } %>
<%-name%>
E. Robert Tisdale wrote:

Can you any provide more details?
Thanks, Steve
<% if( /^image/.test(type) ){ %>
<% } %>
<%-name%>

## Site Timeline

• ### Call for Tutorials @ The 11th International Conference on Computational Creativity ...

• Share To

Polytechforum.com is a website by engineers for engineers. It is not affiliated with any of manufacturers or vendors discussed here. All logos and trade names are the property of their respective owners.