Researchers have constructed a comprehensive open-source ontology designed to represent and organize knowledge within the field of automated planning. Published in the journal Discover Data, the study introduces a structured framework that captures the relationships between planning domains, problems, plans, and the performance of various planners. The primary objective of this work is to leverage the vast amount of empirical data available from planning competitions to enable more efficient reasoning about planner capabilities and domain requirements. By formalizing these concepts, the team aims to facilitate tasks such as automated planner selection and the generation of human-understandable explanations for generated plans.
Automated planning is a sub-field of artificial intelligence focused on finding a sequence of actions that transitions an agent from an initial state to a desired goal state. This process is formally defined as a tuple consisting of states, actions, a transition function, an initial state, and a goal state. In classical planning, actions typically have unit cost and take unit time, with state information modeled using predicates. The field features numerous planners, each combining specific search algorithms and heuristics to generate plans of varying quality, cost, and optimality. The International Planning Competition serves as a critical benchmark for evaluating these systems across a wide range of domains, providing rich empirical data on planner performance.
Traditionally, improving planner performance has relied on manually exploring and curating combinations of algorithms, heuristics, and configurations to identify optimal setups. While effective, this approach is time-consuming and lacks scalability in a field with an ever-expanding range of tools and domains. The absence of a structured framework to represent this information has limited the ability to reason efficiently about planner suitability for specific tasks. Previous efforts, such as the PLANET ontology, offered limited support for reasoning about planner performance and were often closed-source, hindering their reuse and extension by the broader research community.
To address these gaps, the authors developed a planning ontology based on supported competency questions that cover different aspects of automated planning. These questions address planner capabilities, domain requirements, problem structure, plan evaluation, and explanation generation. The ontology is composed of 19 distinct classes and 25 object properties, formulated using Description Logic to provide a formal framework for reasoning. It is designed to be data-agnostic, meaning it models planning concepts independently of any specific competition year, allowing it to ingest newer benchmarks and planner results with minimal effort.
The ontology is segmented into four main categories: Domain, Problem, Plan, and Planner. The Domain category captures characteristics such as requirements, predicates, actions, preconditions, effects, constants, and types. The Problem category represents specific instances within a domain, including objects, initial states, and goal states. The Plan category, adapted from the P-Plan ontology, captures the sequence of actions and their associated costs, which are quantified as non-negative integers. The Planner category classifies planners by type and links them to the specific domain requirements they can solve, categorizing their relevance to a domain as low, medium, or high based on the percentage of problems solved.
The researchers demonstrated the utility of the ontology through two practical use cases. The first use case involved identifying the most promising planner for a specific planning domain. Using data from the optimal track of the International Planning Competition, the team assigned relevance levels to planners based on their historical performance. They compared two policies for assigning planners to problem instances: a Random Policy, which selects a planner at random, and an Ontology Policy, which selects the planner with high relevance for the domain from the ontology populated with IPC-2011 data.
In the evaluation, the Ontology Policy outperformed the Random Policy in terms of the average number of nodes expanded to find a solution. The Random Policy failed to solve problems in several domains, including parking, floortile, and tidybot, highlighting the limitations of random selection. The ontology-driven approach allows for the automated selection of planners by leveraging semantic relationships, such as the hasHighRelevance property. This method provides a transparent and efficient way to identify suitable planners without the need for exhaustive manual experimentation.
The second use case focused on explanation generation, addressing the challenge of making automated planning systems more transparent and accessible. The ontology supports the generation of diverse explanation types, including high-level plan summaries, natural language descriptions of plan steps, and justifications for specific actions. By encoding planning domain knowledge and action semantics within the ontology, the system can extract contextually rich explanations through SPARQL queries. This approach bridges the gap between machine efficiency and human understanding, facilitating better human-AI collaboration.
To assess the effectiveness of the generated explanations, the team conducted a preliminary user survey involving graduate students and researchers in computer science. Participants were presented with problem instances from the Blocksworld and Sokoban domains, along with ontology-generated explanations for three categories: action-level, predicate-level, and plan-level. The survey results indicated that the Plan Explanation category received the highest average score of 4.20 on a 5-point Likert scale, suggesting it was the most clear and helpful. Explaining Specific Actions followed with a score of 3.93, while Explaining Non-Selection of Specific Actions received a score of 3.73.
The study acknowledges several limitations. The effectiveness of the use cases depends on the completeness and representativeness of the data within the ontology. If the data is incomplete or disproportionately represents certain domains or planners, the performance of the ontology-driven reasoning may be limited. Additionally, the template-based explanations, while useful, may not meet all users’ needs for personalized and context-sensitive responses. The authors plan to conduct a more extensive survey in the future to further validate their findings and explore the integration of the ontology with Large Language Models for enhanced reasoning and explanation capabilities.
The ontology and associated resources have been made available to the AI and data communities to promote further research. The project adheres to FAIR principles, ensuring that the data is Findable, Accessible, Interoperable, and Reusable. The ontology is accessible through a persistent URL and a public GitHub repository, which contains ontology model files, mapping scripts, and utility scripts. This open approach facilitates community-driven modifications and periodic updates, ensuring the ontology remains aligned with evolving standards and practices in the field of automated planning.
By providing a structured representation of planning knowledge, this ontology offers a valuable resource for researchers and practitioners in automated planning. It simplifies planning tasks and boosts efficiency by enabling rigorous comparisons of planner performance across different domains. As the field of AI planning continues to evolve, the integration of ontology-based reasoning with other advanced techniques holds promise for advancing the state-of-the-art. The work demonstrates how semantic web technologies can be effectively applied to organize complex empirical data, supporting both automated decision-making and human-understandable explanations in AI systems.
Subject of Research: Development of a comprehensive planning ontology for automated AI planning knowledge representation and application.
Article Title: Building a planning ontology to represent and exploit planning knowledge and its applications
Article References: Muppasani, B. C., Gupta, N., Pallagani, V., Srivastava, B., Mutharaju, R., Huhns, M. N., & Narayanan, V. (2025). Building a planning ontology to represent and exploit planning knowledge and its applications. Discover Data, 3(1), Article 55. https://doi.org/10.1007/s44248-025-00093-9
Image Credits: AI Generated
DOI: 10.1007/s44248-025-00093-9
Keywords: automated planning, ontology, artificial intelligence, planner selection, explanation generation, knowledge representation, semantic web, International Planning Competition, Building, planning, represent, exploit
News Source: Denise Maddox. (November 9, 2026). New Ontology Maps AI Planning Knowledge for Better Planner Selection. Scienmag.
