Online Testing with Reinforcement Learning

Online testing is a practical technique where test derivation and test execution are combined into a single algorithm. In this paper we describe a new online testing algorithm that optimizes the choice of test actions using Reinforcement Learning (RL) techniques. This provides an advantage in covering system behaviors in less time than with a purely random choice of test actions. Online testing with conformance checking is modeled as a 1.5 player game, or Markov Decision Process (MDP), between the tester as one player and the implementation under test (IUT) as the opponent. Our approach has been implemented in C#, and benchmark results are presented in the paper. The specifications that generate the tests are written as model programs in any .NET language such as C# or VB.