Parallel Backpropagation Neural Network
Neucessor
Posts: 7
Hello everybody. I'm going to make a neural network for obstacle avoidance on a robot.
The training algorithm will be made parallel to utilize the multiple cogs.
Would be grateful for any suggestions.
Thanks in advance.
The training algorithm will be made parallel to utilize the multiple cogs.
Would be grateful for any suggestions.
Thanks in advance.
Comments
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
JMH
Post Edited (James Michael Huselton) : 2/25/2009 11:42:59 AM GMT
I'm involved with the robotics research at my university, and we're still trying to break through the learning curve of the propeller board.
However, we're currently working with the intelliibrain robot, and have just completed the communications and navigation portion. We plan on implementing neural networks
into our grid solving algorithms, but haven't started on it yet.
I'd love to see what you come up with, and I will post my findings when we get started.
I cannot stress enough that this is the correct tool for serious robotics research with the P8X32. One month is free, cost less than a bag of weed to own.
Mail me again wheny you understand the ViewScope. Hanno speaks very fluent English, and is very helpful if you email him. We worked for over a year to bring this project to market - let's get going!
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
JMH
Planning to train the net offline (with C on a PC), get the weight constants result, then putting it in the Propeller. Therefore the Propeller will act as a neural processor.
The parallelism technique may either be: training set, pipelining, by neuron (output), or by synapse(input); right now I'm yet to decide which one is most suitable for the problem, I'm going to try from the simplest one first [noparse];)[/noparse]
Btw, this is a bachelors final project, and will possibly be used in a mobile rescue/firefighting robot contest.
The input vector is from 5 ultrasonic sensors. Left, Right, Front, Slight Left/Right (45 degrees). The distance reading will need to be below a certain threshold value to indicate the robot is too close to a wall/obstacle. The output vector will either be in movement states, or (proportional PWM) motor output with direction.
The motor driver used is L298, motor is EMG30. Made own development board so all I/O pins can be accessed (also have the Demo Board).
Next development after this project is complete is to use the CMUCam3 for vision: object recognition.
I will keep posting significant progress.
Cheers
That might be a project in itself?
Jim-
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Signature space for rent, only $1.
Send cash and signature to CannibalRobotics.
Regards,
Xander
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
| To know recursion, you must first know recursion.
| Current projects and ramblings: I'd Rather Be Building Robots
I promise to share all of my research on this forum, once I get the release forms.
▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
JMH
Post Edited (James Michael Huselton) : 2/28/2009 2:39:19 AM GMT
He used the I2C bus protocol, there weren't any major problems. At a distance from candle flame the (eight) arrays will have very different readings (one case is because only an array coincidentally points right to the flame). As it gets closer to the heat source, the readings become more similar across the array. This is all I know right now, hope it helps.
I successfully interfaced the SRF05 ultrasonic sensor (for measuring distance). It worked perfectly well measuring pulse width input using the counter mode POS Detector, but the code was written in Spin instead of assembly. Is this alright?