Mohamed Oubbati, Paul Levi, and Michael Schanz (2005) "A FIXED-WEIGHT RNN DYNAMIC CONTROLLER FOR MULTIPLE MOBILE ROBOTS." Proceedings of the 24th IASTED International Conference MODELLING, IDENTIFICATION, AND CONTROL
==================================================================================================

"We demonstrate the ability of a single fixedweight RNN to act as a dynamic controller ... The controller is properly trained to exhibit adaptive behaviour after its weights have been fixed."

--------------------------------------------------------------------------------------------------
In the control domain, it is shown in [3] that a RNN can be trained to act as a stabilizing controller for three unrelated systems and to handle switch between them. This capability is acquired through prior training; instead of learning data from one system, the network was able to learn from several different systems. A theory explanation about this property can be found in [4][5].

     [3] Feldkamp, L. A., & Puskorius, G. V, (1997) "Fixed weight controller for multiple systems." IEEE Int. Conf. on NN, Vol 2, Texas, USA 9-12 June 1997, 773-778. => done

     [4] Andrew.D. Back (1997) "Multiple and time-varying dynamic modelling capabilities of recurrent neural networks." Neural Networks for Signal Processing 7, IEEE Press, 1997. => done

     [5] Andrew D. Back, Tianping Chen (2002) "Universal Approximation of Multiple Nonlinear Operators by Neural Networks." Neural Computation 14(11), 2561-2566. => done


--------------------------------------------------------------------------------------------------
The adaptive behaviour of RNNs with fixed weights is named differently. It is termed 'meta-learning' in [6], and 'accommodative' in [7]. Such 'multiple modelling' capabilities of RNNs are potentially useful in mobile robotics control where fast adaptive behaviour of the controller is required.

     [6] Prokhorov, D., Feldkamp, L., and I. Tyukin (2002) "Adaptive Behavior with Fixed Weights in Recurrent Neural Networks: An Overview." Int. Joint Conference on Neural Networks, Honolulu, Hawaii, May 2002. => done

     [7] J. Lo (2001) "Adaptive vs. Accommodative Neural networks for Adaptive System Identification." Int. Joint Conf. on Neural Networks, 2001, 2001-2006. => done

--------------------------------------------------------------------------------------------------
In our previous work [8], a novel RNN called Echo State Network (ESN) is used to develop a dynamic controller for mobile robots, and implemented successfully on a real omnidirectional robot.

[8] M. Oubbati, P. Levi, M. Schanz (2004) "Recurrent Neural Network for Wheeled Mobile Robot Control." WSEAS Transaction on Systems, vol. 3, August 2004, 2460-2467. => none



Recurrent-Neural-Network-for-Wheeled-Mobile-Robot-Control

--------------------------------------------------------------------------------------------------
Adaptation here is the ability of the resulting fixed-weight ESN to recognize the robot parameters variations only through its inputs, and to adjust its behaviour to these changes, without changing any synaptic weight.

--------------------------------------------------------------------------------------------------
Echo State Network is a novel RNN in a form of gDynamic Reservoirh(DR), which contains a large number of sparsely interconnected neurons with non-trainable weights. ESN has an easy training algorithm, where only the reservoir-tooutput weight connections are to be adjusted using a linear MSE minimization.