Automated Controlled Experimentation on Software by Evolutionary Bandit Optimization.

Controlled experiments, also called A/B tests or split tests, are used in software engineering to improve products by evaluating variants with user data. By parameterizing software systems, multivariate experiments can be performed automatically and in large scale, in this way, controlled experimentation is formulated as an optimization problem. Using genetic algorithms for automated experimentation requires repetitions to evaluate a variant, since the fitness function is noisy. We propose to combine genetic algorithms with bandit optimization to optimize where repetitions are evaluated, instead of uniform sampling. We setup a simulation environment that allows us to evaluate the solution, and see that it leads to increased fitness, population diversity, and rewards, compared to only genetic algorithms.