12.07.2015 Views

Wireless Ad Hoc and Sensor Networks

Wireless Ad Hoc and Sensor Networks

Wireless Ad Hoc and Sensor Networks

SHOW MORE
SHOW LESS
  • No tags were found...

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

Congestion Control in ATM <strong>Networks</strong> <strong>and</strong> the Internet 93Z −1Z −1Z −1Z −1X(k − 5)X(k − 4)X(k − 3)X(k − 2)X(k − 1)V TW TZ −1 ˆ ( )InputX(k)Purelin Tansig PurelinFIGURE 3.4Two-layer neural network architecture for traffic approximation.is a scalar value, which gives the approximated value of the traffic at eachswitch. The size of the weight matrix for Vk is 6 by 6 <strong>and</strong> for Wk ˆ ( ) it isa 6 by 1 vector at each switch. The activation function, φ(), ⋅ used in theinput <strong>and</strong> the output layers at each node is linear, whereas a tangentsigmoid function is used at each node in the hidden layer. Each switchwill have its own NN congestion controller.The feedback gain, k v , is selected as [0.1 0] T. The initial adaptationgain α, is taken as 0.9 <strong>and</strong> is updated using the projection algorithm asα ( ) 01 .k =.TThe parameter Γ is chosen to be 0.001 as per the( 01 . + φ( xk ( )) φ( xk ( )))requirement, <strong>and</strong> the buffer occupancy is considered empty initially. Theinitial weights are all chosen as zero for the case of NN with no offlinetraining. Here, one can select any initial values for the weights. In this paper,a comparison is shown between NN, with <strong>and</strong> without a training phase.Using 4000 frames of data, which is r<strong>and</strong>omly chosen from a Star Warsmovie, the NN is trained by the well-known backpropagation algorithm(Jagannathan 2006). We saved the weights after training <strong>and</strong> deployedthem as initial weights for the online training. The weight-tuning updatesderived therein were deployed for online training. The network model<strong>and</strong> the traffic sources used in this work are discussed next.3.3.3.2 Network Model <strong>and</strong> Performance MetricsThe network model used in simulations is similar to that shown inFigure 3.2 with several ATM switches in t<strong>and</strong>em. The maximum buffer

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!