From Stories to Cities to Games: A Qualitative Evaluation of Behaviour Planning

Mustafa F. Abdelwahed, Joan Espasa, Alice Toniolo, Ian P. Gent · 2026

Read From Stories to Cities to Games: A Qualitative Evaluation of Behaviour Planning by Mustafa F. Abdelwahed, Joan Espasa, Alice Toniolo, Ian P. Gent (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-04911.txt.

Introduction

Mustafa F. Abdelwahed and 3 co-authors evaluate behaviour planning, a diverse planning approach, across stories, cities, and games in a 2026 paper. Diverse planning aims to produce a set of plans that genuinely differ from one another, useful in risk management, stream analysis, and malware detection. Behaviour planning extends older methods by building a diversity model directly into planning and supporting several planning categories.

The authors test the approach qualitatively in real-world-style settings rather than only on abstract benchmarks. Showing one framework serving narrative, urban, and game domains argues for its generality. Planners that return varied options, not near-copies of one answer, give decision makers real choices when the first plan fails or circumstances shift mid-execution.

Catalog record

CREATORMustafa F. Abdelwahed, Joan Espasa, Alice Toniolo, Ian P. Gent
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE29 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-04911

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

From Stories to Cities to Games: A Qualitative Evaluation of Behaviour Planning Mustafa F. Abdelwahed, Joan Espasa, Alice Toniolo, Ian P. Gent Abstract The primary objective of a diverse planning approach is to generate a set of plans that are distinct from one another. Such an approach is applied in a variety of real-world domains, including risk management, automated stream data analysis, and malware detection. More recently, a novel diverse planning paradigm, referred to as behaviour planning, has been proposed. This approach extends earlier methods by explicitly incorporating a diversity model into the planning process and supporting multiple planning categories. In this paper, we demonstrate the usefulness of behaviour planning in real-world settings by presenting three case studies. The first case study focuses on storytelling, the second addresses urban planning, and the third examines game evaluation. K eywords Diverse Automated Planning \cdot Applications 1 Introduction Classical planners focus on generating a single solution, referred to as a plan. In contrast, diverse planners produce multiple plans that are distinct from one another for a given task. There are several reasons why a user may require multiple plans. While 12 showed that generating several different solutions can help account for future situations, 18 demonstrated that generating different plans is beneficial because such plans may account for side information (e.g., user preferences) that hard to model in the problem. In practice, the generated solution can sometimes be challenging to implement; thus, offering a set of alternative solutions is often more useful, enabling users to choose among different viable options [ 14 , 8 ] . From an application viewpoint, various real-world applications use diverse planning approaches as a core component. For instance, 23 applied diverse planning to anticipate a range of potential future scenarios from a risk management perspective. In this setting, diverse planners are used to generate and rank multiple expected future scenarios. Another important application domain is malware detection [ 4 , 25 ] , where diverse planners are employed to identify malicious activity within network streams. In this context, a planner attempts to generate plans that explain observed network behaviour in order to detect attacks. Using diverse planning techniques enables the detection of multiple potential malicious activities rather than focusing on a single explanation. Beyond these domains, diverse planning also underpins several other applications, including pipeline generation and learning in machine learning [ 15 ] and business process automation [ 6 ] , further highlighting its broad applicability. Several diverse planning frameworks have been proposed to generate k optimal or sub-optimal diverse plans [ 26 , 21 , 28 ] . The most recent approach is proposed by 1 and is referred to as behaviour planning . In this approach, diversity is represented using behaviour spaces , a concept inherited from the Quality-Diversity Optimisation field [ 16 ] . A behaviour space is an n-dimensional grid in which each dimension corresponds to a feature of interest to the user, and each cell within the grid is referred to as a behaviour . The diversity planning problem is formulated as the task of finding a set of plans that maximises the number of distinct behaviours. Behaviour planning comprises two primary components: the Behaviour Sorts Suite (BSS) and Forbid Behaviour Iterative ( \texttt{FBI}_{\texttt{}} ). The former is a qualitative framework used to describe the diversity model, while \texttt{FBI}_{\texttt{}} is a planning approach that uses the diversity model defined by BSS to generate diverse plans. Two implementations of behaviour planning are available. The first [ 1 ] is based on planning-as-satisfiability and targets planning problems modelled using declarative languages such as PDDL (model-based); this implementation is referred to as \texttt{FBI}_{\texttt{SMT}} . The second [ 2 ] targets planning problems that are modelled using simulators and does not require an explicit domain model (model-free); this implementation is referred to as \texttt{FBI}_{\texttt{LTL}} . In this paper, we present three real-world case studies that demonstrate the benefits of behaviour planning. The first case study explores the use of behaviour planning for generating diverse narratives in storytelling [ 5 ] . In this setting, we employ \texttt{FBI}_{\texttt{SMT}} to generate a set of distinct story plots, where diversity is defined in terms of the values of a preselected set of fluents in the goal state. The second case study focuses on urban planning and investigates the generation of diverse plans for this domain [ 3 ] . Here, we apply \texttt{FBI}_{\texttt{LTL}} to produce alternative layouts for the Town of St Andrews, with plan diversity characterised by factors such as sustainability and diversity. The final case study delves into the concept of game replayability [ 17 ] . In this case study, we employ \texttt{FBI}_{\texttt{LTL}} to assess the replayability of the classic Super Mario Land (Game Boy version) by generating a variety of actions that lead to winning the 1-1 world. We distinguish between two types of plans for this case study: those that involve Super Mario killing or avoiding an enemy (e.g., Gomba). These benefits are expressed from a user’s viewpoint. 1 employed three personas proposed by 7 to represent users for behavior planning. These personas were: (i) the end user, who interacts with the planning system through a user interface; (ii) the domain expert, who sets high-level mission objectives for the planning system; and (iii) the algorithm designer, who generates plans based on the end user and domain designer’s requirements. In this work, we will cover each case study’s end user separately. Regarding the domain expert, we consider ourselves as the domain expert, and the algorithm designer will be 1 and 2 since we are using their implementations. The remainder of the paper is organised as follows. We first provide background on the diversity planning problem and behaviour planning. We then present the three case studies in detail. Finally, we conclude with a discussion of limitations, possible improvements, and directions for future work. Figure 1: Three diverse narratives for the Aladdin world. 2 Background This section provides the necessary background for our approach. We first define the classical planning problem, then introduce the diversity model used to represent and quantify plan diversity, and finally describe the behaviour planning framework that generates diverse plan sets. 2.1 Planning Problem Inspired by 9 , we define the planning problem as follows. Definition 1 (Planning Problem) A planning problem is a tuple \Xi=\langle S,A,\gamma,\operatorname{cost},I,G\rangle , where: • S is a finite set of states • A is a finite set of actions • \gamma:S\times A\rightarrow S is a transition function that maps each state s\in S and action a\in A to a successor state \gamma(s,a)=s^{\prime} • \operatorname{cost}:A\rightarrow\mathbb{R}^{+} is a cost function assigning a non-negative cost to each action • I\in S is the initial state, and • G is a formula representing the goal condition. A plan \pi is a sequence of actions \pi=\langle a_{1},a_{2},\ldots,a_{m}\rangle such that each a_{i}\in A and executing \pi from I yields a state satisfying G ; that is, \gamma(\ldots\gamma(\gamma(I,a_{1}),a_{2})\ldots,a_{m})\models G . We denote by \Pi_{\Xi} the set of all valid plans for a given planning problem \Xi . The cost of a plan is defined as the sum of its action costs: \operatorname{cost}(\pi)=\sum_{i=1}^{m}\operatorname{cost}(a_{i}) . While our formulation includes the cost function for generality, we note that

Read the complete text