OCR Text |
Show Figures 5a and 5b are the direction history plots for this particular run at 100% and 70% load, respectively. In both cases, search begins in the 1530 direction, reflects off of a stability limit, and proceeds to another edge of stability. These plots are very telling, for in both cases the global optimum condition is not reached. The system essentially gets hung up on a local peak, or can not find the direction in which to continue searching (which would be exactly along the stability limit line). This demonstrates two very important weaknesses in the direction-set technique: First, if the controller reaches a local peak, it will remain there; second, if a path of ascent exists from a point, but it is narrow, the algorithm may take a long time finding that path. ~ ..; < CI) CI) C1) u >< ~ 0.20 0.15 0.10 0.05 0.00 0.6 0.65 0.7 0.75 0.8 0.85 Swirl Intensity, S' 0.9 0.95 0.70 to 0.72 0.67 to 0.70 0.65 to 0.67 0.63 to 0.65 0.60 to 0.63 0.58 to 0.60 0.56 to 0.58 0.53 to 0.56 0.51 to 0.53 0.49 to 0.51 0.46 to 0.49 1 Figure Sa. Direction history for the counter-swirl nozzle, at 100% load. The performance history results (absolute and average) for a typical genetic algorithm search with a change in load is shown in Figure 6. Note the jagged nature again, which is typical of the genetic algorithm's character. It's difficult to tell much from this plot, other than performance of the burner has been improved over time, and this improvement is repeated following a change in load. 10 |