Behaviour planning

: a feature-based diversity planning approach

Abstract

Background: Diversity planning is applied across a range of real-world applications, including risk management, automated data-stream analysis, and malware detection. Existing formulations typically represent diversity using a distance function, which is computationally inexpensive but limits the level of diversity detail that can be modelled. More importantly, such a representation provides limited explanatory capabilities, as it collapses multiple aspects of diversity into a single numerical function.

Objective: This thesis introduces a new formulation of the diversity planning problem that enables more expressive and interpretable modelling of diversity. The formulation uses an n-dimensional grid, where each dimension corresponds to a user-defined feature. This change shifts the representation of diversity from a metric-based to a feature-based approach. Building on this, the thesis presents a novel diversity planning framework, called Behaviour Planning, that generates diverse plans directly using these customisable diversity models.

Methods: We develop two implementations for this framework. The first is a model-based approach using planning as satisfiability, suitable for problems modelled in declarative languages. The second is a model-free approach that uses tree search and is designed for planning problems whose dynamics are captured by simulators.

Results: Empirical evaluation shows that behaviour planning generates significantly more diverse plans than existing diversity planning methods, while enabling users to define richer, more interpretable diversity models. Additionally, the implementations are the first diverse planners to support planning categories beyond classic planning tasks for model-based problems, and the first diverse planner to support model-free problems. In summary, the results establish that explicit, feature-based diversity modelling leads to more interpretable plan sets.

Future Work: Drawing on the thesis findings, we outline several new research directions, each formulated as an additional research question.

Date of Award	1 Dec 2026
Original language	English
Awarding Institution	University of St Andrews
Supervisor	Ian Gent (Supervisor), Alice Toniolo (Supervisor) & Joan Espasa Arxer (Supervisor)