4 We did try .simpler” search techniques, such as manual trialand-error or automatic hill-climbing with random restarts and momentum, but the end result was significantly inferior. We believe GA offers a good trade-off between simplicity and effectiveness in this context.
5 The applications that we use for training are fft,mg,and radix (Section 5.2). We picked these because they are the fastest to simulate among the parallel applications that we evaluate. By using a small subset, picked not based on behavior but simply on execution time, we speed up training and at the same time minimize the chance of overfitting the final solution to our application set.