A Probabilistic Semantics for Defeasible Argumentation in Systems Validation

Abstract

Engineers commonly validate complex systems by considering arguments, pro and con, that a system possesses (or will possess) desired properties, such as safety or high-quality performance. This argumentation is often conducted in service of a decision, such as passing a critical design review or deploying a system for a given task. It can question facts, methods, and inferences, including generalizations from test cases, analogies between application tasks, and expert opinion. However, very few formalisms support quantitative reasoning about system properties in the presence of such arguments. This article develops a representation for stating conflicting arguments and evaluating such structures into probabilistic assertions about claims. We employ Toulmin argumentation models as a base, an informal representation of rhetorical argument that connects premises through warrants (justifications) to claims. We provide a novel, probabilistic interpretation of warrants as methods for making defeasible inferences by separating the relevance (or aptness) of the method from the likelihood of its conclusion. We use this interpretation to formally integrate Toulmin models into influence diagrams. We illustrate the value of this integration for systems validation by employing the machinery of decision analysis to examine a decision in the presence of engineering arguments. We consider a choice between two rover wheel designs, given a simulation package that affirms one option, and an expert who casts doubt on the validity of the simulation. Then, we determine the expected value of conducting a physical test to resolve a key uncertainty in the presence of these conflicting arguments.