Lose the Order, Keep the Hierarchy: Deordering HTN Plans

Hierarchical Task Network (HTN) planning is a method used in artificial intelligence to solve complex problems by breaking them down into smaller, manageable tasks. While much of the research in this field focuses on how to generate these plans, there has been little work on optimizing them once they are created. This paper addresses this gap by introducing techniques to "deorder" HTN plans—a process that removes unnecessary constraints on the sequence of actions while ensuring the plan remains valid and functional.

The Challenge of Plan Deordering

In planning, a "deordering" process identifies which actions in a sequence do not actually need to happen in a specific order to achieve the desired goal. By removing these rigid constraints, the resulting plan becomes more flexible. While this has been studied extensively in classical planning, it is more complex in the HTN setting because the actions are tied to a hierarchical structure of task decomposition. The authors aim to bring these optimization benefits to HTN planning by adapting existing classical methods to respect these unique hierarchical rules. The same reasoning question is explored in Naive Prompt Optimization, which adds a research perspective.

Adapting Classical Techniques

The researchers adapted two established plan deordering techniques from classical planning. The core challenge was extending these methods so they could account for hierarchical decomposition constraints. By integrating these constraints, the researchers ensured that the deordering process would not break the logic of the hierarchical plan. This allows the system to strip away redundant ordering requirements that were originally imposed during the plan generation phase, resulting in a more streamlined and flexible sequence of actions.

Evaluating Performance

To test their approach, the authors evaluated their methods using the IPC 2023 Partial-Order HTN benchmarks. They compared their results against Optiplan, an existing HTN planner that is designed to generate partially ordered plans directly. The results showed a substantial reduction in the number of ordering constraints in both of the implemented techniques. The same ai evaluation question is explored in From Deceptive Outputs to Deceptive Mechanisms, which adds a research perspective.

Key Findings and Limitations

The study demonstrates that it is possible to significantly simplify HTN plans after they have been generated. While the researchers also observed a reduction in the "critical path length"—the minimum time required to complete the plan—they noted that these improvements were less pronounced than the reduction in ordering constraints. This suggests that while the plans become much more flexible, the overall duration of the tasks may not always decrease as dramatically. The same reasoning question is explored in Spurious Advantage Hidden in GRPO, which adds a research perspective. as detailed in the full paper on Arxiv