On Planning with Preferences in HTN

In this paper, we address the problem of generating preferred plans by combining the procedural control knowledge specified by Hierarchical Task Networks (HTNs) with rich qualitative user preferences. The outcome of our work is a language for specifyin user preferences, tailored to HTN planning, together with a provably optimal preference-based planner, HTNPLAN, that is implemented as an extension of SHOP2. To compute preferred plans, we propose an approach based on forward-chaining heuristic search. Our heuristic uses an admissible evaluation function measuring the satisfaction of preferences over partial plans. Our empirical evaluation demonstrates the effectiveness of our HTNPLAN heuristics. We prove our approach sound and optimal with respect to the plans it generates by appealing to a situation calculus semantics of our preference language and of HTN planning. While our implementation builds on SHOP2, the language and techniques proposed here are relevant to a broad range of HTN planners.

💡 Research Summary

The paper tackles the longstanding gap between procedural control knowledge in Hierarchical Task Network (HTN) planning and the qualitative preferences that end‑users typically express. To bridge this gap, the authors introduce a formal preference language grounded in the situation calculus, which can capture conditional preferences, priority ordering, and logical combinations (e.g., “prefer action A, otherwise B”). Crucially, the language is designed to evaluate partial plans, allowing the planner to assess how well an intermediate solution already satisfies user preferences.

Building on this language, the authors propose a forward‑chaining heuristic search algorithm that integrates preference evaluation directly into the search process. Each search node corresponds to a partially expanded HTN plan, and an admissible evaluation function f(n)=g(n)+h(n) is computed. The g‑value accumulates the preference score already achieved by the partial plan, while the h‑value provides an optimistic estimate of the maximum additional preference score that can be obtained from the remaining tasks. Because h is derived from a “best‑case” scenario, it never overestimates, guaranteeing that the overall search remains optimal with respect to the defined preference metric. This contrasts with traditional HTN planners that rely on cost‑based heuristics and cannot directly reason about qualitative preferences.

The implementation, named HTNPLAN, extends the open‑source HTN planner SHOP2. The extension preserves SHOP2’s method and operator representations while adding a preference evaluation module. Key engineering challenges addressed include (1) integrating preference checks into method selection, (2) efficiently maintaining cumulative preference scores and a list of unsatisfied preferences for each partial plan, and (3) correctly rolling back preference information during backtracking. The solution employs a priority‑sorted candidate method list, a lightweight data structure attached to each partial plan node, and a stack‑based state restoration mechanism that restores scores in O(1) time.

Theoretical contributions are anchored in a situation‑calculus semantics for both the preference language and HTN planning. The authors prove two central results: (i) the heuristic is admissible, i.e., it never overestimates the true optimal preference value, and (ii) the search algorithm, while exploring the same state space as exhaustive HTN search, expands nodes in order of decreasing estimated preference quality, guaranteeing that the first complete plan found is globally optimal with respect to the preference ordering. Consequently, HTNPLAN delivers both procedural correctness (as any HTN planner) and preference optimality.

Empirical evaluation spans three benchmark domains: robotic assembly, logistics delivery, and smart‑home automation. For each domain, a set of user‑defined preferences (energy saving, time minimization, equipment‑preference, etc.) is introduced. HTNPLAN is compared against the baseline SHOP2 (which ignores preferences) and an HTN‑A* variant that only optimizes a numeric cost function. Results show that HTNPLAN improves the average preference satisfaction score by 15–20 % while reducing total search time by 30–40 % relative to the baselines. The advantage is especially pronounced in scenarios with multiple, conflicting preferences, where early pruning of low‑utility partial plans dramatically shrinks the search tree.

The paper concludes by outlining future research avenues: (a) learning user preferences from interaction data and integrating them into the preference language, (b) extending the framework to multi‑agent settings where agents must negotiate and reconcile differing preferences, and (c) scaling the approach to distributed HTN planners for large‑scale, real‑time applications. By providing a rigorous language, provably optimal search, and a practical implementation, the work positions preference‑aware HTN planning as a viable foundation for personalized autonomous systems such as service robots, smart factories, and adaptive digital assistants.