The objective of this research is to implement an artificial neural network into a closed-loop model based control law that requires a state space model. Model based control laws are strongly preferred due to the existence of mathematical proofs for their performance. With mathematical proof of performance, the controller is guar- anteed to work correctly within a designed domain of operation. However, deriving physics based mathematical models and accurately identifying the parameters of such models are cumbersome, or sometimes impossible, for complex nonlinear systems. This limits the application domain of model based controllers.
Neural networks are a great tool for modeling nonlinear systems, because they eliminate the physics based modeling and ease the identification process. However, controllers that have been introduced for neural network model lack mathematical proof of performance. Without mathematical proof of performance, the domain of operation in which the controller works correctly is unknown. So, unexpected con- troller behaviors may emerge.
The methodology proposed in this thesis combines the ease of modeling us- ing neural networks with the mathematical proof of performance using model based control design.Identification data is collected from the system by applying known inputs and recording the system response. A neural network is fitted to the collected data to generate a neural network model. Using the proposed method in this thesis, a state- space model is extracted from the neural network model. A model based controller uses the extracted state-space model to generate control commands.
In this thesis, a mobile robot is used as an example to demonstrate the im- plementation of the proposed methodology. First, as a proof-of-concept, the control commands are applied to the simulated neural network model representing the mo- bile robot. The results show that the robot successfully executes any user-defined command. Next, the methodology will be tested on a real mobile robot, and needed improvements will be made.
http://www.uah.edu/news/research/students-autonomous-robot-project-could-be-a-lifesaver
Comments
Thanks for that and this was my past research and now searching for full time positions. Actually I was using 2 neurons with 10 hidden layers. Learning phase of the neural net was done using the sliding based control law. The reaction time is 0.05 seconds and response plotted is as expected.
Hi, this looks interesting. How many neurones and layers are you using in your neural net ?
How did you do the Learning phase of the neural net ? What is the reaction/lag time of the neural network ? (do you get good square responses or do you get oscillations ?)