Introduction
This project concerns the responsible adoption of AI-driven decision pipelines in real-world applications. To facilitate responsible AI adoption, we are developing methodologies for learning and quantitative robustness evaluation of AI-driven decisions with robustness guarantees.
Explaining the science
Recent advances in AI have led to AI-driven decision pipelines being widely adopted in a variety of real-world applications such as financial systems. However, these AI-driven decision pipelines often lack rigorous robustness guarantees required for the responsible adoption in safety-critical and sensitive applications. We will focus on classes of parametric models that underpin such AI-driven decision-making processes and formalise the desired concept of robustness, which intuitively corresponds to the stability of outputs/outcomes with respect to the changes of model inputs or parameters. The developed methodology will encompass a modelling framework for learning AI-driven decision pipelines, quantitative robustness metrics (including the impact of decision pipelines on robustness), algorithms for robustness testing and estimation (including robustness to distribution shifts, parameter changes and decision pipelining), optimisation of the models and policies (such as parameter optimisation given an objective function), prototype software implementation (where feasible), and evaluation of the resulting techniques on relevant case studies.
Project aims
- Understanding the challenges of robust and responsible use of AI-driven decisions in the real world.
- Leveraging existing tools to establish a framework for modelling AI-driven decision pipelines and clearly defining the different types of robustness desirable for these pipelines.
- Establishing a framework for quantifying robustness and for comparing the robustness of different models and pipelines.
- Designing algorithms for robust decision policies and decision pipeline optimisation.
Applications
The outcome of the project will benefit:
- Financial and other industries, by providing methodologies for robust AI-driven decision pipelines with rigorous guarantees, robustness evaluation of these pipelines, and relevant use cases.
- The academic community, by formulating the types of robustness relevant in real applications, outlining research challenges, and establishing future research directions.
Recent updates
Sample Efficient Model-free Reinforcement Learning from LTL Specifications with Optimality Guarantees