Partial Grounding in Planning using Small Language Models

F. Areces, B. Ocampo, C. Areces, M. Dom\'inguez, and D. Gnad. Partial Grounding in Planning using Small Language Models. In Proceedings of the 2023 Workshop on Knowledge Engineering for Planning and Scheduling, Prague, Czech Republic, 7 2023.

Download

[pdf] 

Abstract

The aim of classical automated planning is to find a sequence of actions, a plan, that changes the state of the world from a given initial state to a state that satisfies the goal condition. Most research in the field focuses on heuristic search, which attempts to find a plan on a fully grounded model of the planning task. However, obtaining the full grounding is often infeasible as its size can be exponentially larger than the original input. We follow up on previous work that introduced partial grounding for planning using a relevance prediction estimate obtained from classical machine learning models. These models are trained offline, on a per-domain basis, to estimate how likely it is for a plan to include a given action. In this article we leverage recent advances in the field of language models in natural language processing (NLP) to improve these estimates. We use small language models to create word embeddings for actions and facts directly from their textual representation. These models provide fixed-length representations for actions and facts reached along a deleterelaxed solution of a planning task, which can be obtained efficiently. We show that these feature vectors can be used to train predictors of action relevance, that consistently identify relevant actions on an established set of hard-to-ground planning benchmarks.

BibTeX

@InProceedings{areces:keps23,
  author =       "F. Areces and B. Ocampo and C. Areces and
                  M. Dom{\'i}nguez and D. Gnad",
  title =        "Partial Grounding in Planning using Small Language Models",
  booktitle =    "Proceedings of the 2023 Workshop on Knowledge Engineering for Planning and Scheduling",
  year =         "2023",
  abstract =     "The aim of classical automated planning is to find a
                  sequence of actions, a plan, that changes the state
                  of the world from a given initial state to a state
                  that satisfies the goal condition.  Most research in
                  the field focuses on heuristic search, which
                  attempts to find a plan on a fully grounded model of
                  the planning task. However, obtaining the full
                  grounding is often infeasible as its size can be
                  exponentially larger than the original input. We
                  follow up on previous work that introduced partial
                  grounding for planning using a relevance prediction
                  estimate obtained from classical machine learning
                  models. These models are trained offline, on a
                  per-domain basis, to estimate how likely it is for a
                  plan to include a given action.  In this article we
                  leverage recent advances in the field of language
                  models in natural language processing (NLP) to
                  improve these estimates. We use small language
                  models to create word embeddings for actions and
                  facts directly from their textual
                  representation. These models provide fixed-length
                  representations for actions and facts reached along
                  a deleterelaxed solution of a planning task, which
                  can be obtained efficiently. We show that these
                  feature vectors can be used to train predictors of
                  action relevance, that consistently identify
                  relevant actions on an established set of
                  hard-to-ground planning benchmarks.",
  address = "Prague, Czech Republic",
  month = "7",
}