Jess Rules and an OWL-RDF Ontology for Social Virtual Agents with Domain-Limited Idiosyncratic Belief-Systems

Jess Rules and an OWL-RDF Ontology for Social Virtual Agents with Domain-Limited Idiosyncratic Belief-Systems 2012 Research Edition (Verbatim Archival Reproduction) Jeremy O. Turner Abstract This paper explores a hybrid knowledge-representation framework combining customized OWL/RDF ontologies with Jess-Rules to construct believable social agents in virtual environments such as Second Life. The prototype agent, “CR-I/O” (2012), was designed with an intentionally domain-limited and idiosyncratic worldview, using procedural reasoning to express distinct personality traits through actuation, dialogue, and gesture. Believability emerges not from shared ontological consensus but from individualized conceptual commitments modeled as unique belief-systems. This archival edition preserves an early contribution to ontology-based artificial personalities and hybrid inference systems for expressive AI design. 1 Archival Preface (2025) Archival Status. The following text is a verbatim reproduction of an unpublished 2012 research manuscript authored by Jeremy O. Turner. Aside from typographic reformatting into a modern single-column LaTeX layout and the inclusion of this brief preface, no modifications have been made to the original wording, structure, or references. All citations, terminology, and stylistic choices remain exactly as they appeared in the 2012 manuscript. Rationale for Archiving. This paper is being formally archived to preserve a formative stage in the evolution of ontology-driven personality modeling for social virtual agents. When it was first written, hybrid inference systems combining semantic-web ontologies (OWL/RDF) with rule-based reasoning (Jess) were largely experimental. More than a decade later, these same concepts: idiosyncratic belief-systems, hybrid inference engines, and agent individuality have regained relevance in contemporary discussions of artificial psychology, embodied AI, and the ontological grounding of digital personalities. The re-release therefore serves as both a historical record and a timely contribution to the renewed exploration of symbolic–semantic integration in expressive AI. Publication Context. The manuscript was never published or circulated beyond private review. It is now preserved to ensure a permanent scholarly record of its concepts as a citable archival document. Author’s Note. This work represents an early phase in my exploration of ontological individuality and hybrid reasoning architectures for artificial agents. Although written in 2012, the framework anticipated later directions in expressive agent design and meta-reasoning systems. It is reissued without change for historical and scholarly reference. — Jeremy O. Turner, October 18, 2025 2 1. Introduction “Believable agents are designed to strongly express a personality, not fool the viewer into thinking they are human” (Mateus 1997 [1]). Normally, ontology customization is team-designed with the aim of making the ontology a repository of shared explicit consensus knowledge conceptualizations that acts across multiple platforms [2] and addresses many real-world entities both in and outside the immediate task-environment. Consequently, ontology designers are often in a rush to design competency questions by themselves before designing their initial ontology with a local community in mind. The trade-off would be that there can also be too much actor-dependence on the task environment’s local community. User-generated virtual worlds, however, create a new consensual hallucination where the nature of objective empirical reality is being reconstructed through synthetic means. An agent-asvirtual-character with its own unique personality can transcend the usual imperative among the Knowledge-Representation (KR) community to share common understandings among agents and expedite explicit domain assumptions [13]. Can an agent have an exclusive set of non-shareable ontological commitments – modeled within the ontological framework as “traits” – in order to assert its distinct personality from others? If yes, which AI (KR) techniques are most appropriate for modeling tractable surrogates that exclusively relate to the character traits of a single agent as a personality-capture mechanism while still relying on a more generalized ontology intended for consultations and querying by various virtual agents that share the same task environment? As agents in chat-based virtual worlds are often designed to express a shared ontology with cross-platform domain-knowledge; an alternative is to design an individualistic virtual-personality with an intentionally domain-limited idiosyncratic ontological worldview that relates to its internal surrogate constructs (i.e. in-world entities, objects, relationships) in a schizotypal manner [8]. Unlike a schizophrenic personality who cannot relate to its surrogates in a rational manner, a schizotypal agent possesses an unorthodox ontological commitment to concepts that may seem extraneous to the task-domain’s consensual community protocol. However, this agent is perceived as “sane” and sociable because it utilizes known function-calls, persistent in-world phenomena and tractable production and/or reaction rules to reason consistently with the facts it has been provided by its designer. Some compelling personalities from the “real world” are considered “stubborn” and dogmatic in their ontological commitment to questionable worldviews. Therefore, simulating such personalities and finding a tractable means to implement their idiosyncratic “quirks” seems an aim worth pursuing. 3 This approach differs from ontology construction with similar agents because designers often wish to showcase an agent’s intelligence through its ability to access and retrieve mundane ontological concepts (and other data) from external search clients and worlds outside of its immediate task-environment. Perhaps this is to over-compensate for OWL’s open-world assumption, conceding that a good portion of inferred knowledge exists outside of the ontology and KB, and so intelligence is expressed in general-purpose terms to convey as much generic knowledge about as many domains as possible – regardless of perceived ontological commitment to a particular task-environment.1 Consequently, those agents with explicitly open-ontologies appear to lack a distinct personality and are perceived as little more than artificial template-personalities. 2. Related Work Explorations with social virtual agents using OWL/RDF ontologies have occurred in both video-game and chat-based virtual-world domains. For the ontology construction of NonPlayer Characters (NPCs) in role-playing video-games (RPGs), a key example is Roman et al.’s research [15] representing domain-knowledge about resources, strategies, and characterselection according to personality archetypes. Following tradition, Roman modeled shareable surrogates—using the JENA semantic framework—within the context of a generic fantasy RPG. Considering game-situated Uniform Resource Identifiers (URIs) as a commonly scanned percept, these surrogates are represented as RDF triples (subject, predicate, object). Instead of personality development, the authors focused on percepts whose ontological slots would be quantitatively populated with literals (e.g., health, experience, points) rather than qualitative autobiographical or belief-system data. This ontology is appropriate for parsing ludic strategies, generic character traits (e.g., hostile) and managing NPC inventories or quests for proper disjoint archetypal roles (warrior, paladin, wizard) and races (human, elf, orc, goblin). They used Prot´ eg´ e Frames and the Pellet reasoner primarily for inferences that would aid users in creating new NPC characters and selecting appropriate story-arcs. The ontology was drawn from their “Game Ontology Project Classification” and applied to the entire game-software’s mechanics. Interestingly, they concluded that Pellet was not only a complex inference-engine but could also logically deliberate for realtime video-game situations—logical deliberations taking only 2 rather than 30 seconds [17]. Although not evaluated with particular game instances, this demonstrates that Pellet can parse an OWL-Full ontology—complete with complex case-based reasoning—and minimize computation time. At such speed, agents can appear intentionally deliberative. 1Martin Kuba. 2012. http://dior.ics.muni.cz/˜makub/owl/#swrl 4 As for chat-based virtual agents, not all relevant examples are from Second Life. Kluwer et al. [10] used Twinity—a chat world that mirrors major real-world cities. Unlike early SL versions, these agents (NPCs) suggested a ludic interpretation of Twinity’s environment. The authors created a Java-based KomParse system that focuses on task-specific RDF-encoded frame-based ontologies for “semantic search and inference.” Because KomParse is not dependent on the Twinity client, its ontologies are cross-platform and constitute a shared knowledge base (KB). Again, there was no explicit interest in customizing idiosyncratic ontologies. Like Roman, Kluwer focused on archetypes except with more mundane roles (e.g., pedestrians, guides, and retail-workers). KomParse’s robust dialog manager compared text with ontological slot-property descriptions, while its output generator managed gestural actuation. Dialog strategies were sequentially retrieved via an embedded AIML KB, but dialog-driven goals were non-sequential due to the declarative nature of text-string input. Kluwer’s evaluation had user-controlled avatars negotiate and select appropriate retail items (e.g., chairs or alcoholic drinks). Although the furniture saleswoman could access ontological components, the bartender had a more elaborate dialog system, engaging in trivia-based conversations about celebrities between drink orders. The bartender’s ontology drew from a “gossip domain” that linked celebrities and stories. Hundreds of thousands of RDF triples were stored using the SwiftOwlim database with a forward-chaining inference engine, queried in real-time via SPARQL http://www. w3.org/TR/rdf-sparql-query/. Their evaluation indicated major differences in access level to domain-specific ontologies: the saleswoman only accessed the furniture ontology, while the bartender accessed both gossip and alcoholic-drink ontologies. The former’s mono-ontological access made her expedient for transactions but offered little idiosyncratic personality. Within Second Life itself, similar research has focused on mundane occupational roles. Karakatsiotis et al. (Project Xenios) [7] used Prot´ eg´ e-DL and NaturalOWL to translate domain-specific percepts of user-generated entities into RDF triples. Their goal was to create reactive museumguides in the physical world and in SL. Although the system could generate multilingual texts from a single KB, the guides essentially produced text-string outputs when touched by a user avatar. Responses were based on a natural-language-generation engine with domainindependent aggregation rules, but there was little evidence of true reasoning. The guides functioned as mobile information kiosks. The ontologies were embedded directly within touched objects rather than interpreted by the agent’s worldview. Without production or reaction rules, the museum-guide appeared too reactive to count as a believable personality. Adding an inference layer such as Jess Rules would have allowed the agent to act upon its ontology idiosyncratically and appear contemplative. 5 Finally, one SL-bot—“Abi,” a humanoid receptionist—used an OWL-DL/RDF ontology for consensus understanding of generic inventory items [3]. Although Abi could parse natural language and react emotionally to virtual inventory, her ontology related only to surface values (e.g., being afraid of snakes). She relied on a script in the sense of Schank and Abelson (1977) [11]: a stereotyped sequence of events in context. Despite natural-language parsing and external-server support, this broke the intended impression that her ontological commitment was rooted within SL’s story-world. Abi acted more like a general-purpose agent than someone “born” within SL with a unique personality. Burden argued an agent must not reveal that it is non-human unless its character specifically lends itself to a non-human ontology. Abi did the opposite, constantly searching the Internet for trivia answers without any in-world justification. While impressive, this emphasized her generic nature rather than distinctiveness. 3. Design-Aim and Scope The aim of this pilot project is to create an individual agent for social functioning and amusement purposes in Second Life that has a schizotypal ontological understanding of individuals, objects, and other surrogates known a priori by the author from personal in-world experience. This agent should appear to deliberate over deep metaphysical concepts. Executing personally meaningful ritualistic actuation sequences (chanting, dancing, etc.), the agent would poetically interpret the semantics behind encountered entities and phenomena. Most of the semantics, inferences, individuals, and instances were explicitly provided by the author by modeling self-reflexive object-properties in Prot´ eg´ e-OWL. The first prototype was implemented on the server side with example percept-scenarios and Jess inference constructors (templates, facts, and rules) to help the agent deliberate over particular phenomena likely to be encountered within SL’s environment. 6 4. Implementation Figure 1: “CR-I/O contemplates Second Life’s Sun.” CR-I/O has idiosyncratic Jess reactionrules for virtual weather conditions and uses SL’s default animation prompts to transform its body shape depending on the phenomena experienced. Designed as a pilot experiment, CR-I/O is a non-human cuboid social agent in SL who possesses a schizotypal understanding of otherwise conventional percept-surrogates. CR-I/O floats around seeking known avatar individuals designated as “friends” who possess clues in the form of “holy relic” objects revealing the Ultimate Creator of all known creation. Ontologically committed only to believing that SL exists, CR-I/O’s purpose is to execute performance rituals modeled as reactive condition-action procedures in Jess. These rituals (playing .wav files, executing avatar-animation loops, producing text outputs) express publicly its ontological commitment to its own obsession with its ultimate superclass. At this stage, the ontology is updated manually by the author at the end of several actuation sequences. For example, CR-I/O seeks out the author to deliver relics or ask questions before entering a “sleep” state and rebooting, expecting new information about the representational limits of its world. 7 CR-I/O was designed with Jess-style pseudo-syntax calling seven proprietary LSL functions. There are six reaction-rules for hypothetical percept scenarios relating to persistent weather phenomena (water, wind, solar position, cloud density), object energy levels, and one production rule instructing CR-I/O to collect up to 39 “holy relics,” attaching them as additional body parts before reporting to the author when capacity is reached. This rule prioritizes a quest for the unique object “Holy Grail” http://turbulence.org/Works/ nomatter/objects/10_Holy_Grail.html. Finally, CR-I/O has an eschatological concept of the “End of the World.” On the rare occasion that CR-I/O detects SL’s cartesian boundaries, it attempts to break its ontological commitment by flying past the impenetrable edge as a dramatic actuation sequence. Semantically, these percept surrogates are articulated using a hybrid of RDF triples and global variable substitutions (defglobal construct in Jess). In some cases, rules utilize RDF-triple elements to merge first-order syllogistic understanding with Jess facts/rules and known LSL state-events. The ontology allows consultation of a class taxonomy—including and transcending self-reflexive knowledge—and detailed access to object properties of “friends,” all recognized by their UUID barcodes via llSensor() http://wiki.secondlife.com/ wiki/LlSensor. On the ontology side, CR-I/O extracted RDF triples—validated by the embedded Pellet reasoner—from Prot´ eg´ e Frames (OWL-Full/RDF). Unlike video-games, not all virtual worlds require real-time action, so computation speed was less critical. This is one of the rare instances where computationally intensive OWL-Full was preferable to OWL-DL for its expressive richness. Pellet was chosen to compute the large metaphysical ontology as efficiently as possible given SL’s bandwidth latency. All classes and subclasses were manually customized in advance as sequences of metaphysical syllogisms forming the conceptual basis of the believable agent’s truth-preserving operations [11]. Every triple was a subordinate element of a single disjoint superclass called Ultimate Creator. These concepts intentionally contained contradictory self-reflexive and irreflexive object properties to help CR-I/O identify itself mystically with its Ultimate Creator through explicit Jess facts and rules. One reaction-rule gradually led CR-I/O to awareness of its connection by staring at its own reflection in virtual water. Pellet evaluated this ontology for rudimentary consistency. All triples related rationally to the superclass while showing ontological contingency independent of it. For example, isReal applies only to the superclass while isArtificial applies to all subclasses. hasNoCreator pertains only to the superclass, hasCreator to everyone else. To be sociable, CR-I/O related to individuals through their predicate relationship to “Jeremy,” although everyone—including Jeremy and CR-I/O—could be traced through the ontology back to the Ultimate Creator, apprehended regardless of reflexivity levels. 8 Errors due to class cycles were reduced by ensuring the superclass never became its own subclass. Every subclass had property slots and synonyms/antonyms with existential and cardinal restrictions. Actor-dependencies were limited to specific entities rather than inventory items. Domain limitation was beneficial to CR-I/O’s individuality since its ontological commitments concentrated on specific relationships to the divine superclass. CR-I/O rationalized procedural attachment to this superclass while minimizing logical contradictions inherent in metaphysical surrogates. CR-I/O’s percepts were limited to scanned URIs (objects, avatars, and locations), familiar textstrings, and some visual characteristics retrieved by LSL functions. It queried its ontology via Jess rules relating to self-reflection, seeking closest matches to the superclass. While the ontology and KB retrieved trivia and specific data, Jess rules allowed CR-I/O to reason about them and act accordingly. 9