MK

Research

My work is in AI planning, also known as automated planning, with a focus on the methodologies and systems needed to make planning work in practice. I often address aspects that receive little attention in planning-oriented venues such as ICAPS, with the aim of connecting theoretical advances to use-inspired research, in the sense of Stokes’ Pasteur’s Quadrant.

AI Planning

Engineering AI Planning Systems

AI planning systems need to be designed, built, and composed to meet specific requirements. This theme covers the software development life cycle for planning systems, the functionalities that planning systems consist of, toolboxes of planning services, provenance, and the experience of practitioners. One of the guiding questions is whether tailored combinations of granular planning functionalities can solve complex problems better than off-the-shelf planners.

Planning for Real-World Problems

Real-world problems place demands on AI planning, such as complex constraints and risk-aware decision-making. The focus is on what makes planning domains realistic and how these aspects can be integrated into planning, in particular Hierarchical Task Network (HTN) planning.

Neurosymbolic Planning

Symbolic planning can be combined with learning-based methods, in particular Large Language Models (LLMs) and reasoning models. This theme examines how such models can support the engineering of planning domain models, from their generation and refinement to their configuration and validation, how action models can be learned from data such as Internet of Things (IoT) data, and how well these models perform in planning tasks themselves. It also covers learning to guide the planning process, such as method selection in HTN planning.

Green AI Planning

Most of the planning literature optimises planning time. AI planning is, however, not only a solution to energy problems but also a consumer of energy. This theme addresses how much energy planning consumes across its life cycle, from domain modelling and validation to plan generation, which factors determine this consumption, such as planner design, domain model design, and the planning phases involved, and how planning can be made more energy-efficient.

Applications

Application domains include energy smart buildings, autonomous driving, robotics, distributed applications, and activity recognition.

Beyond AI Planning

Automated Service Composition

Service composition shares its structure with planning: services must be found, selected, and composed dynamically into adaptive systems. Current work examines automated service composition using planning and large language models, including Large Reasoning and Action Models and LLM-based agents.

Green ICT

Energy efficiency is a concern for information and communication technology as a whole, not only for AI. This theme contributes to a reference framework for Green ICT.

Learning Algorithms from Data

Machine learning can be used to derive efficient, domain-specific algorithms from data. This theme develops a framework for learning algorithms and studies the formal properties of the learned algorithms.

Collaboration

If you are interested in collaborating, please get in touch. I do not work on projects in the military domain.