Abstract
dc:description.abstractArtificial autonomous agents are increasingly involved in human society, giving rise to hybrid multiagent systems that integrate both human and artificial actors.<br/>The governance of such hybrid multiagent systems has thus become a pressing and timely concern. In this thesis, I adopt a formal approach to investigate the theoretical foundations of responsibility and norms in multiagent interactions, aiming to leverage these concepts to address governance challenges in such hybrid systems.<br/><br/>Responsibility attribution is a well-established mechanism in the governance of human societies. Formalising the concept of responsibility in alignment with human intuition facilitates its integration into the computation-based decision-making frameworks of artificial agents. In this thesis, I formalise a novel form of responsibility using a modal logic approach, addressing the absence of a notion of active responsibility that is suitable for third-party governance in multiagent systems. Building on this formalisation, I prove the conceptual independence of this new notion from counterfactual responsibility, a widely acknowledged notion in both academic literature and legal practice. Its applicability to artificial agents is further demonstrated by the development of model-checking algorithms and an analysis of their computational complexity.<br/><br/>The responsibility gap is one of the major concerns in multiagent system governance. It refers to the situation in which an outcome happens while no agent can be held responsible. In this thesis, I prove that responsibility gaps may exist even when both the proposed notion of responsibility and the standard notion of counterfactual responsibility are taken into account. To address this issue, I explore two approaches: the first involves higher-order responsibility, that is, the responsibility for a responsibility gap; the second involves the use of norms to constrain agents' behaviour. It is proven that the first approach can fill responsibility gaps if the considered order is sufficiently high. The second approach can also achieve this, although the synthesis of such norms suffers from computational intractability. To tackle this challenge, I propose approximation algorithms for norm synthesis that achieve nearly optimal approximation factors.<br/><br/>Once norms (whether or not synthesised to achieve responsibility gap-freeness) are imposed by an external governor, agents within the system are expected to interpret and reason about those norms. To formalise this, I adopt a deontic-logical approach and propose a modal language with four modalities to capture four forms of agentive permission statements. Mutual undefinability results demonstrate the semantic independence of these permission notions. Model-checking algorithms and computational complexity analysis further illustrate their effectiveness in practical applications. To support agents' reasoning about permissions, I further propose a sound and strongly complete axiomatisation.<br/><br/>From a conceptual perspective, this thesis engages with some of the most intuitive ideas underlying human society regulation and explores how they can be adapted to the context of AI-integrated hybrid societies. From a technical perspective, it draws on formal tools and techniques from multiple research disciplines and helps bridge problems in the broader field of AI governance. Overall, this thesis contributes to the formal understanding of responsibility and norms in multiagent systems, bridging intuitive human concepts with computational frameworks. By combining insights from logic, ethics, and algorithmic design, it offers a principled foundation for addressing pressing governance challenges in AI-integrated societies.
Degree
thesis:*- Name dc:type.qualificationname
- Ph.D.
- Level dc:type.qualificationlevel
- doctoral
- Grantor dc:publisher.institution
- University of Southampton
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Shi, Qi
- Advisors dc:contributor.advisor
-
- Naumov, Pavel
- Cirstea, Corina
- Ozbek, Kemal
- ORCID person:identifier.orcid
Rights
- Language dc:language
- English, eng
Identifiers
dc:identifier.*- DOI dc:identifier.doi
- https://doi.org/10.5258/SOTON/01/T138