Towards Value Alignment for Opaque Agents Through Concept Analysis and Inter-Agent Value Modelling

Value alignment, ensuring that artificial intelligence acts in ways aligned with humans, is a key challenge in the deployment of AI systems and an important aspect in the development of hybrid intelligence. A notable issue in value alignment is the lack of precision around what value alignment actually is and how to evaluate its success. Another issue is the fact that multiple agents are involved in value alignment, but these agents' values may not be easily understood or expressed. In order for autonomous agents to readily adapt to these agents' needs, and to promote interpretability of the system, an evidence-based means to assess the value priorities of agents is required. In my PhD, I am proposing a precise definition of value alignment through a qualitative concept analysis, synthesising themes from across the literature to produce a cohesive description. In addition, I am developing a novel mechanism for inter-agent value inference to enable agents to assess other agents' values in an interpretable way. My PhD will generate a cohesive description of the value alignment problem, identifying sub-problems for further research, and build a means for autonomous agents to perform value inference in pursuit of achieving value alignment.