Wireless Ad Hoc and Sensor Networks

104 Wireless Ad Hoc and Sensor NetworksMoreover, it is observed that the two-layer NN method performed betterin providing a low CLR during congestion, compared to the adaptiveARMAX, threshold, and one-layer methods. However, as expected, themultilayer NN method takes longer to transfer cells from the source tothe destination compared to the open-loop scenario. In addition, providingoffline training did indeed reduce the cell-transfer delay, whereas theCLR is still near zero. From the figures, it is clear that the CLR resultingfrom the two-layer NN method outperforms the adaptive, thresholding,and one-layer NN methods. Finally, the overall delay observed for twolayerNN was within 1% of the total transmission time of 167 sec, whereasother schemes took longer than 3%.Figure 3.15 and Figure 3.16 present the buffer utilization using adaptiveARMAX, one-layer NN, multilayer NN with and without a priori training.The buffer utilization is very low for the thresholding method because ofa threshold value of 40%, whereas the one-layer NN stores cells in thebuffer frequently resulting in a very high utilization. The multilayer NNwithout a priori training does not utilize the buffer as much as the multilayerNN with training. This is due to the inaccurate knowledge of thetraffic flow with no a priori training. Also, as more buffer space is providedto the multilayer NN with no a priori training, the utilization increaseswith buffer size. However, as pointed out earlier, the overall delay issmaller than other schemes even though the queuing delay is slightlyhigher.100Feedback delay = 0Buffer utilization9080706050ThresholdAdaptiveOne layerTwo layer w/o trainingTwo layer with training4030150 200 250 300 350Buffer length (cells)FIGURE 3.15Cell loss ratio with congestion.