ORLA: Learning Explainable Argumentation Models

This paper presents ORLA (Online Reinforcement Learning Argumentation), a new approach for learning explainable symbolic argumentation models through direct exploration of the world. ORLA takes a set of expert arguments that promote some action in the world, and uses reinforcement learning to determine which of those arguments are the most effective for performing a task by maximizing a performance score. Thus, ORLA learns a preference ranking over the expert arguments such that the resulting value-based argumentation framework (VAF) can be used as a reasoning engine to select actions for performing the task. Although model-extraction methods exist that extract a VAF by mimicking the behavior of some non-symbolic model (e.g., a neural network), these extracted models are only approximations to their non-symbolic counterparts, which may result in both a performance loss and non-faithful explanations. Conversely, ORLA learns a VAF through direct interaction with the world (online learning), thus producing faithful explanations without sacrificing performance. This paper uses the Keepaway world as a case study and shows that models trained using ORLA not only perform better than those extracted from non-symbolic models but are also more robust. Moreover, ORLA is evaluated as a strategy discovery tool, finding a better solution than the expert strategy proposed by a related study.