Sign In

Consolidating LAMA and Best-First Width Search for Efficient Agile Planning

Core Concepts
Combining the strengths of the LAMA and Best-First Width Search (BFWS) planning algorithms can lead to a new state-of-the-art agile planner that outperforms both individual approaches.
The paper explores the combination of the LAMA and Best-First Width Search (BFWS) planning algorithms to achieve improved performance in agile planning, where the goal is to find plans as quickly as possible without considering plan quality. The authors first attempt to simply add the strongest open-list used in BFWS(f6) to LAMA, but find that this combination actually harms performance compared to the individual planners. However, a detailed ablation study reveals that selectively removing certain features from LAMA and BFWS can lead to significant improvements. The authors propose a new planner called NOLAN that consolidates the open-lists from LAMA and the novelty-based open-list from BFWS(f hLM 2 ). NOLAN outperforms all other state-of-the-art agile planners, including the winners of the previous International Planning Competitions (IPCs), by a large margin, achieving a 13% higher agile score than the second-best method. The key insights are: Combining the full set of open-lists from LAMA and BFWS(f6) is not beneficial, but selectively removing certain features leads to significant improvements. Using the hLM heuristic in the novelty-based open-list is consistently superior to using hFF. Removing the open-lists containing only preferred operators (hFF and hLM ) from LAMA yields the highest coverage, while removing the hLM open-list improves expansions and agile scores. The resulting NOLAN planner establishes a new state of the art for agile planning, outperforming all other recent planners.
The number of expanded states for NOLAN is 1644.0. The agile score for NOLAN is 1756.3.

Key Insights Distilled From

by Augu... at 04-30-2024
Consolidating LAMA with Best-First Width Search

Deeper Inquiries

How could the memory bottleneck for VisitAll tasks be addressed in NOLAN?

The memory bottleneck for VisitAll tasks in NOLAN could be addressed by implementing a strategy similar to what BFWS-based planners use. When the estimated memory usage for storing the seen fact tuples exceeds a certain threshold, NOLAN could fall back to using a simpler strategy, such as setting the novelty value to a constant like 1. By implementing this fallback mechanism, NOLAN can avoid running out of memory on tasks with a large number of facts and different heuristic values, such as VisitAll tasks.

What other exploration techniques could be combined with NOLAN to further improve its performance?

To further improve the performance of NOLAN, it could be beneficial to combine it with other exploration techniques such as Type-based Exploration with Multiple Search Queues. This technique involves maintaining multiple search queues based on different types of information, allowing for a more diverse exploration of the search space. By integrating Type-based Exploration with NOLAN, the planner can benefit from a wider range of search strategies and potentially discover more efficient paths to solutions.

How do the insights from consolidating LAMA and BFWS apply to other planning domains beyond agile planning?

The insights gained from consolidating LAMA and BFWS can be applied to other planning domains beyond agile planning by emphasizing the importance of balancing exploration and exploitation in heuristic search algorithms. By carefully selecting and combining different open-lists, partition functions, and tiebreaking strategies, planners in various domains can achieve a better trade-off between exploring new states and exploiting promising ones. Additionally, the idea of simplifying complex planning strategies to improve performance can be generalized to different problem domains, leading to more efficient and effective planning algorithms overall.