Figure 2 illustrates how COAL deploys a regression model for DSE. As the first phase,
COAL labels a number of randomly generated design configurations by the simulator
(simulation cluster), which constructs the initial labeled data set. At each learning
phase of COAL, the semisupervised learning is utilized to select and label two unlabeled
design configurations. After that, COAL updates each regression tree by the
configuration labeled by the other tree. Using the updated trees, COAL invokes the
active learning engine, via which COAL finds and simulates the unlabeled design
configurations on which the trees show the largest disagreement. This newly labeled
design configuration is then used to further update both trees. Such a procedure is
repeated for several times until the predefined iteration is reached. Benefited from
the collaboration of cotraining semisupervised learning and active learning, COAL
provides us accurate predictions of processer responses at the cost of simulating only
a few design configurations.