Modeling Intelligent System Execution as State Transition Diagrams to Support Debugging

Currently, few tools are available for assisting developers with debugging intelligent systems. Because these systems rely heavily on context dependent knowledge and sometimes stochastic decision making, replicating problematic performance may be difficult. Consequently, we adopt a statistical approach to modeling behavior as the basis for identifying potential causes of failure. This paper describes an algorithm for constructing state transition models of system behavior from execution traces. The algorithm is the latest in a family of statistics based algorithms for modelling system execution called Dependency Detection. We present preliminary accuracy results for the algorithm on synthetically generated data and an example of its use in debugging a neural network controller for a race car simulator. 1 Introduction Traditional approaches to debugging may be inadequate or inappropriate for AI systems. The applications tend to be poorly understood. The base technology (often language..