To demonstrate the effectiveness of COAL, we compare it with the ANN DSE ap proach proposed by I¨pek et al. [2006], and a supervised M5P regression tree. Following the setting utilized by I¨pek et al. [2006], the ANN adopts one 16 unit hidden layer, a
learning rate of 0.001, and a momentum value of 0.5. The minimal number of examples in each leaf of the M5P tree is set to 4. In our experiments, both ANN and M5P models are constructed by a training set consisting of 400 labeled design configurations. Among the 400 labeled configurations, 300 labeled design configurations are the same to the configurations in the initial training set of COAL, and other 100 configurations are generated randomly and labeled by cycle accurate processor simulations. In addition, we also provide the performance data of the intelligent sampling based ANN DSE approach (ANN-IS for short), a variant of the ANN DSE approach proposed by I¨pek et al. [2006], as a reference. Following the setting suggested by I¨pek et al., ANN-IS
repeatedly updates an ensemble of 10 ANNs trained by 10 fold cross validation over the labeled design configurations, and iteratively labels (simulates) the unlabeled configurations on which the ANNs present largest coefficients of variance (ratio of the standard deviation to mean). ANN-IS shares the same initial training set with COAL, and all approaches share the same testing data with COAL.
To demonstrate the effectiveness of COAL, we compare it with the ANN DSE ap proach proposed by I¨pek et al. [2006], and a supervised M5P regression tree. Following the setting utilized by I¨pek et al. [2006], the ANN adopts one 16 unit hidden layer, a
learning rate of 0.001, and a momentum value of 0.5. The minimal number of examples in each leaf of the M5P tree is set to 4. In our experiments, both ANN and M5P models are constructed by a training set consisting of 400 labeled design configurations. Among the 400 labeled configurations, 300 labeled design configurations are the same to the configurations in the initial training set of COAL, and other 100 configurations are generated randomly and labeled by cycle accurate processor simulations. In addition, we also provide the performance data of the intelligent sampling based ANN DSE approach (ANN-IS for short), a variant of the ANN DSE approach proposed by I¨pek et al. [2006], as a reference. Following the setting suggested by I¨pek et al., ANN-IS
repeatedly updates an ensemble of 10 ANNs trained by 10 fold cross validation over the labeled design configurations, and iteratively labels (simulates) the unlabeled configurations on which the ANNs present largest coefficients of variance (ratio of the standard deviation to mean). ANN-IS shares the same initial training set with COAL, and all approaches share the same testing data with COAL.
การแปล กรุณารอสักครู่..