Causal discovery and inference for autonomous-agent interactions

Thesis

Everyday humans interact with others and generally understand how their actions impact them, even when not sharing a common goal. This is particularly crucial in high-risk domains where a failure to comprehend how to react or how others might react could result in injury or death. Thus it is important to develop such functionality for the autonomous agents that increasingly operate around people. This thesis focuses upon the expanding field of autonomous vehicles as agent interactions are prevalent there, yet explainability and transparency are also important. Given the inherently causal nature of agent interactions and that explanations can utilise counterfactual inference this work builds upon causality literature. Four main chapters of content provide the primary contributions of this thesis in addition to a literature review of the aforementioned research areas. The first chapter introduces causality theory and applies it to the autonomous-vehicle domain, before benchmarking causal-discovery methods on real-world autonomous-vehicle data to identify challenges for existing techniques. The second chapter combines an action-based theory of mind and counterfactual inference in order to produce SimCARSv1, which outperforms the methods of the first chapter. The third chapter builds upon the previous by tackling challenges associated with representing a system of interacting autonomous agents via structural causal models. The final chapter concludes by merging the contributions from all previous chapters, along with integrating a means by which to estimate the instantaneous reward parameters of agents. This culminates in the creation of SimCARSv2, which offers similar quantitative performance to SimCARSv1, but with greater expressiveness due to its structural-causal-model-based architecture. The sum of these contributions represents an important step towards bridging the promising fields of causality and autonomous vehicles, with the goal of building autonomous-agent technologies that can interact with humans safely.

Authors: Howard, RPM

• DOI: 10.5287/ora-dayndoxgk
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: causal reasoning; counterfactual inference; theory of mind; autonomous agents; behavioural interactions
• Subjects: Intelligent transportation systems; Autonomous robots; Robotics; Intelligent agents (Computer software); Artificial intelligence; Automated vehicles