Estimating WCET using prediction models to compute fitness function of a genetic algorithm

Genetic algorithms can be used to generate input data in a real-time system that produce the worst-case execution time of a task. While generating the test data, the fitness function is normally evaluated using a cycle-accurate simulator of the processor architecture, which consumes a significant computational effort and time. We propose to replace the simulator-based actual execution with a predictive model that is trained using the samples acquired on the simulator. The feasibility of this proposal was evaluated using four distinct predictive models, namely artificial neural networks, generalized linear regression, gaussian process regression and support vector regression. The results obtained on the four benchmarks Bubble sort, Insertion Sort, Gnome sort and Shaker sorts indicate that the proposed use of prediction models can significantly reduce the temporal verification time. The time gain achieved is up to 17.7 times and the best accuracy achieved is 98.5%.