Introduction
In classical planning, state-of-the-art solvers typically follow a two-stage pipeline: grounding and search. Grounding involves translating a high-level, "lifted" domain description into a concrete, propositional representation. While this process is technically polynomial, it faces a severe "grounding bottleneck." As the number of objects in a domain increases, the number of possible grounded actions explodes exponentially. In complex real-world scenarios, such as large-scale cybersecurity networks or robotics, the grounded task often exceeds memory limits before the search for a plan even begins.
Partial Grounding Framework
To address this bottleneck, we introduce partial grounding, a technique that leverages machine learning to identify and instantiate only the most promising parts of a planning task. Unlike traditional grounding, which aims for reachability, partial grounding treats the process as a selection problem.
The algorithm modifies the standard grounding process through two primary mechanisms:
- Priority-Based Grounding: Actions are instantiated according to a priority queue, where the priority is determined by a learned function estimating the likelihood that an action belongs to an optimal plan.
- Early Stopping: The grounding process terminates once a sufficient portion of the task is instantiated, allowing the solver to proceed to the search phase with a manageable problem size.
Learning Relevance through Relational Context
The core of this approach is the Priority Function. To learn this function effectively across different problem sizes, we utilize relational features that describe an action’s context at a "lifted" level. This allows models trained on small, solvable instances to generalize to massive, previously "un-groundable" instances.
We evaluated several machine learning models, including Logistic Regression (LOGR), Support Vector Regressors (SVR), and Inductive Relational Trees (via the Aleph tool). These models learn rules such as "Is this satellite turning toward a direction required by a goal image?" to filter out irrelevant actions. By learning at the lifted level, the models can be trained on a 10-object problem and successfully guide a 1000-object problem.
Experimental Results and Scalability
The proposed method was tested on standard benchmarks from the International Planning Competition (IPC) and custom large-scale instances. In domains like Agricola, Blocksworld, and Caldera, partial grounding reduced the number of operators by several orders of magnitude compared to full grounding.
As shown in the distributions above, the learned models effectively separate optimal operators from non-optimal ones. This high precision allowed the planner to solve instances with up to 100 blocks in Blocksworld—problems that the baseline full-grounding approach could not even begin to process due to memory exhaustion. Logistic Regression (LOGR) was particularly effective, allowing the planner to find solutions in less than 10 seconds for instances that previously crashed during preprocessing.
Soundness, Completeness, and Future Work
While partial grounding is sound (it only finds valid plans), it is inherently incomplete if the model misses a necessary action. To mitigate this, we implement an incremental loop: if the search fails, the algorithm grounds additional actions and retries.
Future work will focus on more sophisticated stopping conditions and automated feature selection to further reduce the overhead of the training phase. By reframing grounding as a classification task, we open new avenues for scaling classical planning to real-world complexities.
Summary updated