BDI Ontology for Rational Agents

• BDI Ontology is a formal semantic model capturing agents' beliefs, desires, and intentions through a modular, axiomatized design pattern.
• It enables precise agent-based reasoning by integrating logic-augmented LLMs, production-rule engines, and foundational ontologies.
• Empirical evaluations demonstrate enhanced inferencing consistency, efficient RDF serialization, and robust support for multi-agent, neuro-symbolic systems.

The BDI Ontology is a formal knowledge representation designed to capture the cognitive architecture of rational agents via their Beliefs, Desires, and Intentions (BDI), as well as the processes and temporal contexts that govern the evolution of these mental states. It provides a modular and axiomatized ontology design pattern (ODP) aligned with foundational ontologies, enabling semantic precision and interoperability across AI systems, knowledge representation frameworks, and reasoning platforms (Zuppiroli et al., 21 Nov 2025).

1. Formal Structure and Modular Organization

The BDI Ontology is constructed as a set of interlocking modules, each modeled as an ODP and grounded in foundational ontologies such as DOLCE-UltraLite and EventCore. Its architecture encompasses five core modules:

• Core Entities: Models Worlds, Agents, and Mental Entities.
  • Example definitions:
  • $WorldState \sqsubseteq d0:Eventuality$
  • $Agent \sqsubseteq dul:Agent$
  • $MentalEntity \sqsubseteq d0:CognitiveEntity$
• Mental State Module: Defines the principal BDI classes—
  • $$Belief, Desire, Intention \sqsubseteq MentalState \sqsubseteq MentalEntity$$
  • Structural relations (e.g., $Belief \sqsubseteq \forall hasPart.Belief$) and motivational/support links (e.g., $Belief \sqsubseteq \exists motivates.Desire$, $Intention \sqsubseteq \exists fulfills.Desire$).
• Mental Process Module: Represents cognitive activities (e.g., BeliefProcess, IntentionProcess) with process-specific effects on mental states:
  • $$MentalProcess \sqsubseteq d0:Activity \sqcap MentalEntity$$
  • Properties such as $$generates, modifies, suppresses \sqsubseteq affects$$
• Goal and Planning Module: Incorporates $Plan$, $Goal$, and $Task$ hierarchies; situates planning and execution in terms of intentions and procedural decomposition:
  • $Plan \sqsubseteq dul:Plan$, decomposed into $Task$s; $PlanExecution \sqsubseteq dul:PlanExecution$
• Temporal Module: Temporalizes all entities and processes with $TimeInstant$, $TimeInterval$, and associates validity via $atTime$ and $hasValidity$ properties, enforcing constraints such as functional start/end times.

Alignment with foundational ontologies ensures that key classes are sub-classed from, or mapped to, DOLCE-DnS UltraLite (DUL), PROV-O (for provenance), and BasicPlan/Sequence ODPs. This facilitates both rigorous semantic integration and practical reuse in complex systems (Zuppiroli et al., 21 Nov 2025).

2. Formal Semantics and Logical Foundations

The BDI Ontology adopts SRIQ(D) description logic expressivity. Its formal semantics are defined with:

• TBox axioms: Class inclusions and property restrictions (e.g., $Intention \sqsubseteq \exists specifies.Plan$).
• RBox axioms:
  • Role hierarchies: $$generates, modifies, suppresses \sqsubseteq affects$$
  • Inverse properties: $perceives \equiv isPerceivedBy^{-1}$, $isSupportedBy \equiv supports^{-1}$
  • Transitive: $hasPart, follows$
  • Asymmetric: $precedes$ (inverse of $follows$)

Domain and range axioms are declared for all object properties, enforcing semantic integrity and unambiguity in inter-module interactions. Number restrictions are adopted (e.g., $$TimeInterval \sqsubseteq (=1\ hasStartTime.TimeInstant) \sqcap (\leq 1\ hasEndTime.TimeInstant)$$) to ensure proper temporal referencing.

This rigorous axiomatization ensures TBox/RBox-level inferencing is feasible, enabling ontology-based reasoning for both classification and query answering tasks (Zuppiroli et al., 21 Nov 2025).

3. Integration with Logic-Augmented LLM and T2B2T Reasoning

The BDI Ontology is integrated with two complementary reasoning paradigms:

Logic-Augmented Generation (LAG)

By embedding ontology snippets and axioms within LLM prompts (e.g., GPT-4o), the LAG pipeline guides text-to-triples generation:

• Ontology-defined classes/properties constrain RDF/Turtle output.
• Post-processing with an ontology reasoner further enforces ontological validity.
• Example: Contradictory scenarios (“check into hotel at home”) are flagged by the model, which produces both a prevention intention and an alternative plan, demonstrating improved inferential consistency and RDF coherence (Zuppiroli et al., 21 Nov 2025).

Semas Platform: Triples-to-Beliefs-to-Triples (T2B2T)

The Semas reasoning platform implements a production-rule engine:

• Ingests RDF triples, mapping them to $bdi:Belief$ or other mental states.
• Applies procedural rules in Prolog-style: $[Belief]\ /\ [Conditionals] \gg [Actions]$, generating new desires, intentions, or plans.
• Serializes resultant mental state/process instances as RDF, ensuring round-trip semantics for agent belief states and procedural intentions.

This bidirectional paradigm enables agents to synchronize mental states with dynamic knowledge graphs and procedural rule engines (Zuppiroli et al., 21 Nov 2025).

4. Experimental Validation and Operational Capabilities

Empirical evaluation spans two primary scenarios:

• LAG: Over 47 MS-LaTTE dataset cases, ontology-augmented prompts detected more logical contradictions (22% increase in the example test suite) and achieved complete RDF modelling coverage (100% of outputs matching competency questions), compared to ≈85% for unaugmented models.
• Semas T2B2T: In an e-commerce context, agents performed end-to-end intention derivation and RDF serialization with 100% rule coverage and sub-200 ms cycle times, confirming efficiency and scalable integration.

Across both, explicit BDI semantics enhanced consistency, explainability, and interoperability (via a common vocabulary), supporting procedural and declarative system integration (Zuppiroli et al., 21 Nov 2025).

5. Use-Case Patterns and Querying

The BDI Ontology provides a foundation for complex agent-based reasoning and semantic querying:

• Meeting attendance management: Agents ingest schedules, form beliefs and intentions, and decompose goals into plans using the BDI constructs.
• SPARQL and DL Query Support:
  • Retrieval of all intentions and associated plans.
  • DL queries for agents with specific commitments.
  • Justificatory tracing of intention formation.

These capabilities enable cognitively grounded, interoperable multi-agent architectures and seamless integration between Web data, procedural rules, and agent cognition (Zuppiroli et al., 21 Nov 2025).

6. Significance for Multi-Agent and Neuro-Symbolic Systems

The BDI Ontology bridges declarative semantic web formalisms and cognitive-procedural agent architectures:

• Provides an ontological backbone for explainable, interoperable multi-agent systems.
• Enables semantic integration with neuro-symbolic platforms (e.g., logic-augmented LLMs, reasoning engines).
• Supports historical, current, and hypothetical reasoning about agent mental states, plans, and actions.

This formally specified, modular, and interoperable approach facilitates the next generation of explainable, semantically-integrated multi-agent and neuro-symbolic systems, addressing core challenges in both AI and knowledge representation domains (Zuppiroli et al., 21 Nov 2025).