Parallel partitioning for distributed systems using sequential assignment

Predicting Regression Test Failures Using Genetic Algorithm-Selected Dynamic Performance Analysis Metrics

“…As previously indicated, the genetic algorithm allows the Naive Bayes classifier to perform better with larger initial metric volumes (number of metrics used times program code sections the metrics are measured for) than would otherwise be possible. While the GA metric subset selection improvement was less pronounced for the other machine learning algorithms for these small benchmarks, the Genetic Algorithm improvements are already valuable for these benchmarks and are expected to become more pronounced for larger commercial and open source programs as the metric volume increases [22].…”

“…While we restricted our experiments to the well known Siemens test suite [14,15] and eight OpenPAT metrics [2] in this paper, the approach as presented is directly applicable to larger software programs and additional dynamic program analysis metrics. Future work may consider incorporating dynamic metric information gathered during testing (not just prior information gathered for the correct program version) into the method, adding new dynamic measurement metrics from OpenPAT including for example detailed internal control flow information, evaluating the accuracy of the approach with different training sets sizes, different prediction quality metrics, and different code section sizes, and evaluating the benefits of the GA metric selection feature with different machine learning algorithms on larger commercial and open-source programs [22].…”

“…The OpenPAT toolkit includes a wide range of metrics available at a fine-grained level for each code section aka "assembly basic block", in the program. Because a typical program can have a large number of code sections [22], we can quickly obtain an extremely high dimensionality (that being code sections times metrics per section) in the dataset. Unfortunately, many machine learning algorithms do not perform well when data dimensionality is high compared to the number of labelled examples used for training [21].…”