Optimisation in AI (OAI)

Despite rapid advances in AI, many real-world challenges require more than learning patterns from data alone. While modern AI systems excel at extracting insights, recognising complex relationships, and adapting to new information, they often lack explicit representations of objectives, constraints, and system dynamics that are essential for reliable decision-making. Addressing increasingly complex problems therefore requires approaches that combine the strengths of data-driven learning with structured reasoning and optimisation.

The Optimisation in AI (OAI) pillar focuses on developing the foundations and methodologies for adaptive, reliable, and intelligent decision-making. By integrating data-driven learning with structure-driven optimisation, researchers can combine the flexibility of AI with the rigour, interpretability, and verifiability of optimisation and control methods. This enables AI systems to make better decisions in dynamic and uncertain environments, while remaining aligned with real-world objectives and constraints.

By bringing together advances in machine learning, optimisation, and control, this research seeks to tackle problems that cannot be fully addressed by either approach alone, enabling more efficient, trustworthy, and impactful AI solutions for complex real-world challenges.

To address these challenges, we advance new methodologies that unite learning, optimisation, and decision-making. Our research explores how AI systems can reason, plan, and act effectively under real-world objectives and constraints, while remaining adaptive, reliable, and aligned with human needs. Key research areas include:

Research Focus

Integrating foundation models and machine learning with mathematical optimisation for problem formulation, decision-focused learning, solver-in-the-loop reasoning, and human-aligned decision-making.

Developing capabilities that enable AI systems to dynamically formulate, solve, act, and adapt as objectives, constraints, and operating environments evolve.

Establishing efficient and training-free adaption of foundation and generative models, distribution optimisation, sampling, and the mathematical foundations of learning and optimisation.