- The paper introduces a qualitative evaluation of behaviour planning with a novel behaviour space based on Quality-Diversity Optimization.
- It demonstrates the method across storytelling, sustainable urban planning, and game replayability using both model-based and model-free approaches.
- The approach offers actionable insights by analyzing methodological strengths and limitations, paving the way for automated behaviour space design.
Qualitative Evaluation of Behaviour Planning Across Domains
Introduction
This work provides a systematic qualitative evaluation of the behaviour planning paradigm for diverse planning, demonstrating its applicability in narrative generation, urban planning, and game design. Building on prior research in diverse planning that has primarily focused on obtaining multiple distinct solutions to a given planning problem, behaviour planning advances the field by introducing a behavior model grounded in Quality-Diversity Optimization, instantiated as a behaviour space, and by supporting both model-based and model-free planning. The study critically addresses the problem formulation, diversity metrics, and practical impact in each domain, using concrete case studies to elucidate methodological strengths and current limitations.
Theoretical Foundations
Diverse Planning and Behaviour Spaces
The classical automated planning problem is extended from computing a single optimal or feasible action sequence to producing a set of diverse plans. Behaviour planning formalizes diversity through user-specified features over plans, collectively forming a behaviour space. In this n-dimensional grid, each axis corresponds to a feature of interest (e.g., narrative outcomes, urban attributes, player interactions), and each plan is mapped to a cell based on its evaluated feature tuple. The behaviour diversity count is used as an objective, maximizing the number of distinct behaviour-space elements in the returned plan set.
The planning process itself is operationalized by the Forbid Behaviour Iterative (FBIx) algorithm, which iteratively generates a plan, forbids its behaviour from reoccurring, and continues searching for up to k distinct behaviours. The specific implementations used—FBISMT (planning-as-satisfiability) and FBILTL (model-free, simulator-based with LTL constraints)—offer flexibility across domains where declarative models may or may not be tractable.
Case Studies
Storytelling
In the narrative domain, behaviour planning is leveraged to generate plot-diverse story solutions within the Aladdin world, employing the classical planning compilation of intentional narrative models. The feature set captures the possible combinations of "married-to" relationships at the narrative endpoint, constructing a behaviour space where each outcome is differentiated by the specific pairings. FBISMT enables the sampling of plots with distinct narrative outcomes, providing a systematic means for authors to examine space of story trajectories conditioned on high-level goals.
A salient contribution in this case study is the explicit separation between the user interface (plot selection), domain knowledge (goal configuration and feature design), and algorithmic diversity generation. However, the evaluation acknowledges that narrative believability—essential for human-authored stories—is only partially addressed here, as intentionality constraints are weakened by the classical planning transformation. This exposes an open challenge for integrating more sophisticated narrative models without sacrificing the tractability of diversity-driven planning.
Urban Planning
The urban planning case addresses land-use transformation for sustainable city design, extending behaviour planning to cases without explicit goal conditions via horizon planning. The simulator-based, model-free instantiation (FBILTL) is utilized to navigate combinatorial design spaces, where actions correspond to land-type conversions, and the core behaviour features are sustainability and diversity metrics (the latter using Shannon-Weaver entropy).
Plans are considered distinct if they realize different combinations of aggregate scores in the behaviour space. The qualitative demonstration in the town of St Andrews contextualizes how behaviour planning enables urban designers to sample alternative development strategies traceable to trade-offs in sustainability and ecological diversity. The categorical encoding of real-valued sustainability/diversity provides a tractable interface to planning while illustrating the impact of abstraction-granularity in feature space design.
Notably, the approach is compared to Constraint Programming (CP), but is favored due to the explicit execution plans that can be presented to, and selected by, stakeholders, preserving the interpretability of transition paths in urban policy design.
Game Replayability Evaluation
Applying behaviour planning to game analysis, the FBILTL framework is coupled with an emulator to automate the analysis of player trace diversity in Super Mario Land. The behaviour space is defined over features such as enemy engagement, distinguishing between playthroughs where Mario kills or avoids enemies. This method enables a direct evaluation of the replayability of game content, transcending mere procedural content generation by focusing on player interaction diversity.
The case study underscores the modularity of behaviour planning, as demonstrated by the integration of a dedicated A* game planner optimized for Mario agent pathfinding. This approach is shown to facilitate the enumeration of qualitatively distinct player strategies, providing actionable feedback for game designers seeking to quantify and maximize the diversity of gameplay experiences.
Implications and Discussion
Behaviour planning, as evaluated across these three domains, demonstrates robust domain-agnosticism and supports both model-based and model-free paradigms. Distinctively, it incorporates diversity as a primary criterion within the planning loop rather than as an external metric or a post-processing operation. This architectural decision yields richer, user-centered plan sets that enable choice and analysis directly at the behaviour-feature level.
Challenges include the up-front cost of diversity-feature space engineering, which currently relies on domain knowledge and manual feature selection/extraction. There is also an identified gap in supporting more complex narrative and behavioural constraints, particularly in domains where model fidelity and realism (narrative believability, urban regulatory feasibility, advanced game behaviors) are non-trivial to specify.
Future work is oriented towards automating or semi-automating behaviour space construction through learning informative behaviours from data and incorporating interactive refinement processes, thereby scaling behaviour planning to more complex, data-rich, and user-driven settings.
Conclusion
This paper systematically demonstrates the deployment of behaviour planning for generating diverse solutions across storytelling, urban planning, and game design. By integrating explicit behavioural diversity models into the planning process, it enables targeted generation and evaluation of alternative outcomes, addressing both technical and user-centric objectives. Its generality, domain-adaptability, and focus on qualitative plan attributes position behaviour planning as a versatile tool for future AI planning applications requiring structured exploration of behavioural variation.
Reference: "From Stories to Cities to Games: A Qualitative Evaluation of Behaviour Planning" (2601.04911).