A Conceptual Design of Industrial Asset Maintenance System by Autonomous Agents Enhanced with ChatGPT

Example: ACL to NL (Request from AGENT to ChatGPT)
Step 1: AGENT initiates a request for temperature data using ACL.
ACL Request Message:
      • Performative:	request
      • Sender:			AGENT
      • Receiver:		ChatGPT
      • Content:
            (inform
                  :sender 		AGENT
                  :receiver 		ChatGPT
                  :ontology 		"temperature"
                  :content 		(request :content (temperature :location "factory_section_A"))
            )
      • Reply-To:		AGENT
      • Reply-With:		temp_request_01
Translated NL Query (sent to ChatGPT):
"Could you provide the current temperature in factory section A?"
Example: NL to ACL (Response from ChatGPT to AGENT)
Step 2: ChatGPT processes the NL query and responds with a temperature value in NL.
NL Response from ChatGPT:
"The temperature in factory section A is 22 °C."
Step 3: AGENT translates this NL response back into ACL format for structured understanding.
Translated ACL "Inform" Message
      • Performative:	inform
      • Sender:	 		ChatGPT
      • Receiver:	 	AGENT
      • Ontology:	 	temperature
      • Content:
            (inform
                  :sender 		ChatGPT
                  :receiver 		AGENT
                  :ontology 		"temperature"
                  :content		(temperature: location "factory_section_A": value 22: unit "C")
            )
      • In-Reply-To: 	temp_request_01
      • Reply-With: 	temp_response_01

Step 1: AGENT queries for factory equipment maintenance schedules using SPARQL.
SPARQL Query:
     SELECT	 ?maintenanceDate	?equipmentID
     WHERE {
          ?equipment	rdf:type					ex:Equipment.
          ?equipment	ex:location 				"factory_section_B".
          ?equipment 	ex:maintenanceDate 	?maintenanceDate.
          ?equipment 	ex:equipmentID			?equipmentID.
		  }
NL Query (for ChatGPT):
"Could you provide the next maintenance dates for all equipment in factory section B?"
Step 2: ChatGPT responds with the maintenance information in NL.
NL Response from ChatGPT:
"The next maintenance for equipment ID 102 in factory section B is scheduled for 2024-11-15, and for equipment ID 207, it’s on 2024-12-01."
Step 3: AGENT translates the NL response into RDF triples for structured understanding and further action.
Translated RDF Triples:
…
ex:equipment_102 	 rdf:type 			ex:Equipment.
ex:equipment_102 	 ex:location 			"factory_section_B".
ex:equipment_102 	 ex:maintenanceDate "2024-11-15"^^xsd:date.
ex:equipment_102	 ex:equipmentID 	"102" .
    ex:equipment_207  rdf:type 			ex:Equipment.
    ex:equipment_207  ex:location 			"factory_section_B".
    ex:equipment_207  ex:maintenanceDate "2024-12-01"^^xsd:date.
    ex:equipment_207  ex:equipmentID       	  "207".
    …

The OPRA agent architecture represents a hybrid approach combining reactive, deliberative, and knowledge-based systems. It dynamically queries external knowledge sources (such as ChatGPT or similar systems) to address gaps in predefined behavior, distinguishing it from traditional agent architectures that rely on strict rules or internal reasoning mechanisms. ChatGPT, as discussed in [37], serves as an adaptable, up-to-date external resource that complements traditional learning and knowledge systems, making it particularly valuable for real-time decision-making in uncertain environments. This capability aligns with the way OPRA agents can query ChatGPT for external deliberation in dynamic situations. Let us now compare OPRA to other widely discussed architectures in agent theory. 5.1. OPRA vs. Reactive Agent Architectures Reactive agents, like those following the “subsumption architecture” used for behavior-based robotics [38] and its recent implementations [39], focus on direct mappings between observations and actions. They normally do not have internal models or deliberative processes but respond to stimuli based on predefined rules. Reactive agents are very fast and straightforward but lack adaptability to novel situations. In contrast, the OPRA architecture introduces adaptability by dynamically querying external knowledge (ChatGPT) to handle unforeseen or complex situations. Reactive agents struggle with uncertainty because fixed stimulus-response pairs bind them. OPRA, on the other hand, can handle novelty by generating prompts to seek clarifications and guidance from an external model. Reactive agents act without reasoning (or limited, stimulus-driven), while OPRA integrates reasoning capabilities by consulting external knowledge, bringing a semi-autonomous decision-making process. 5.2. ORPA vs. Deliberative Agent Architectures Deliberative agents use symbolic reasoning and planning to decide on actions. The “Belief-Desire-Intention” (BDI) model [40] is a common example, where agents maintain beliefs about the world, set goals (desires), and form plans (intentions) to achieve these goals. BDI agents have internal beliefs and reason about actions using predefined logic or rules. OPRA, in contrast, relies on external knowledge (ChatGPT) to complement its internal state and reasoning. This allows it to address situations where internal knowledge is insufficient. OPRA agents are more flexible in highly dynamic or complex environments where rules are either unavailable or constantly changing. Deliberative agents, while powerful in static, well-understood domains, can become rigid in the face of rapid change or uncertainty. Deliberative agents may suffer from high computational complexity due to internal planning and reasoning. OPRA mitigates this by offloading some reasoning to an external system, making it lightweight in terms of internal decision-making. Recently, in [41], the BDI-Prompting approach was introduced, which incorporates the BDI model into LLMs to enable proactive, goal-directed task planning and provide transparent explanations for agent decisions. While BDI-Prompting and OPRA both focus on enhancing explainability, BDI-Prompting emphasizes internal motivational states—beliefs, desires, and intentions—to guide actions. In contrast, OPRA follows a cyclic observation-prompt-response-action loop, which allows for external feedback and promotes real-time interaction with other agents or users. OPRA’s design facilitates dynamic multi-agent collaboration, where information is exchanged and actions are based on collective intelligence, making it well-suited for complex environments. BDI-Prompting, on the other hand, is more centered on a single agent’s internal decision-making process and does not emphasize the same level of interaction with external agents or ongoing dialogue. This distinction is particularly critical for Industry 4.0 and robotics applications, where ORPA’s adaptive, real-time interaction and multi-agent coordination provide a stronger foundation for autonomous systems operating in dynamic, complex industrial environments. 5.3. OPRA vs. Hybrid (Reactive-Deliberate) Architectures Hybrid architectures combine elements of both reactive and deliberative agents, often using reactive layers for immediate responses and deliberative layers for long-term planning. One example is TouringMachines [42], which incorporates both reactive and deliberative components. Hybrid agents typically separate reactive and deliberative layers. OPRA does not maintain separate layers but rather relies on a feedback loop with an external system (ChatGPT) to augment its reasoning. This makes OPRA more dynamic since it can outsource deliberation rather than maintaining complex planning layers internally. Hybrid architectures are relatively adaptable but can be limited by the complexity of their deliberative layers, while OPRA excels in environments where adaptability is key because it can ask for advice from ChatGPT to handle unexpected or novel situations. While hybrid architectures require careful design to balance reactivity and deliberation, OPRA offloads much of its complexity, allowing for a more scalable solution in dynamic, unpredictable domains like industrial maintenance or healthcare. Xu et al. [43] suggest an interesting recent implementation of hybrid architecture in robotics. In contrast to OPRA, which focuses on a cyclic interaction-driven process (observation, prompt, response, and action), the real-time decision-making model described in [43] adopts a deliberate/reactive hybrid control architecture. While both frameworks aim to handle dynamic, complex environments, OPRA relies on collaborative multi-agent intelligence and interaction with external agents, fostering a continuous feedback loop to adapt to changing conditions. The hybrid architecture in [43], however, combines a deliberate, hierarchical approach to planning with a reactive system for immediate responses to sudden events, such as unknown obstacles. This design is more focused on balancing structured planning with real-time reactivity. OPRA’s prompt-response mechanism is better suited for scenarios where external inputs (from humans or other agents) are crucial, whereas the hybrid architecture excels in environments requiring fast, autonomous responses to dynamic changes. For robotics and Industry 4.0, OPRA’s emphasis on multi-agent collaboration and external feedback offers more adaptability in scenarios requiring human-agent interaction or coordination between multiple systems. 5.4. OPRA vs. Goal-Oriented Architectures Generally, in goal-oriented architectures, agents focus specifically on achieving predefined goals, and their behavior is driven by the success of reaching these goals (see, e.g., [44]). Their decision-making could be deterministic, where the agent selects actions that directly lead to goal achievement without necessarily considering trade-offs between different outcomes. Many of these systems employ hierarchical planning, where complex goals are broken down into sub-goals, enabling the agent to achieve higher-level objectives incrementally. Utility-based agents (see, e.g., [45]) represent a specific subset of goal-oriented architectures. These agents aim to maximize certain utility functions (cost, reward, risk, etc.), often through continuous optimization techniques. They rely on modeling outcomes and weighing various actions to achieve the maximum benefit, frequently making decisions under uncertainty but with known probabilities. When enhanced with game-theoretic insights, utility-based agents are capable of coordination in self-interested multi-agent scenarios [46]. In collaborative scenarios, a shared utility function allows agents to cooperate toward a common goal [47]. While utility-based agents optimize actions based on various reward functions, OPRA differs by not explicitly optimizing for long-term rewards. Instead, OPRA’s cyclic observation-prompt-response-action model leverages external information to guide decision-making when internal knowledge or predefined policies are insufficient. OPRA bypasses the need for complex reward function modeling by querying an external system (e.g., ChatGPT) for advice, allowing the agent to act on context-specific knowledge rather than relying on exploratory learning. This contrasts with goal-oriented architectures that focus solely on predefined goals, whether they are structured hierarchically or through continuous utility optimization. Therefore, OPRA offers a more flexible, adaptive approach, which could be beneficial in dynamic and uncertain environments (particularly suited for real-world applications in Industry 4.0 or robotics) where predefined goals or utility functions are hard to establish. 5.5. OPRA vs. Learning Agents Learning agents, especially those employing ML (e.g., reinforcement learning), improve their behavior over time by interacting with their environment (see, e.g., [48]). These agents focus on optimizing long-term outcomes based on accumulated experience. The deep reinforcement learning framework, as introduced in [49], combines the strengths of deep neural networks and reinforcement learning to achieve efficient decision-making in complex, dynamic environments. This approach is fast becoming dominant in smart manufacturing [50]. Learning agents generally update their policies or models through continuous learning from past interactions. In contrast, OPRA does not focus on learning from experience but seeks external advice based on current conditions, making it more suitable for real-time, dynamic scenarios where learning may be too slow or ineffective. While learning agents need time to adapt through iterative updates, OPRA can instantly adjust by querying ChatGPT in unfamiliar situations, providing a quicker response to novelty. Furthermore, OPRA leverages the broad, generalized knowledge of an external system (ChatGPT), enabling it to address scenarios outside its direct experience, whereas learning agents may struggle to generalize beyond their training data. 5.6. OPRA vs. Knowledge-Based Agents Newell’s work [51] is one of the foundational studies introducing the concept of knowledge-based agents, emphasizing the separation between knowledge and the program executing it. Jackson’s study [52] is widely regarded as a classic reference on expert systems, a key subclass of knowledge-based agents that utilize rule-based inference mechanisms. In their comprehensive textbook, Russell and Norvig [53] further discuss knowledge-based agents, expert systems, and rule-based reasoning. Knowledge-based agents rely on structured, rule-based reasoning applied to a static knowledge base. They use predefined rules and logic to infer conclusions, often excelling in specialized domains like medical diagnosis or fault detection. These agents depend heavily on the quality and comprehensiveness of their knowledge base, which requires continuous maintenance and updates. In contrast, OPRA involves dynamic querying of external systems like ChatGPT, enabling it to remain adaptive in environments where knowledge evolves rapidly, or new challenges arise unexpectedly. Knowledge-based agents require ongoing maintenance to update their knowledge base and inference rules, often involving domain experts to manage this process. OPRA, however, avoids this burden by dynamically querying systems like ChatGPT, which are continuously updated, thus reducing the need for manual updates to the agent’s internal knowledge base. 5.7. Summarizing Unique Characteristics of OPRA Unlike most agent architectures that rely on internal mechanisms (e.g., rules, learning models, planning), OPRA dynamically queries an external knowledge source (such as, e.g., ChatGPT) to address gaps in the agent’s understanding or reasoning capacity, giving OPRA agents the unique ability to consult external expert advice as needed. The OPRA model encourages an iterative process where the agent can ask clarifying questions until a sufficiently confident answer is reached, making it particularly useful in ambiguous, evolving, adversarial, or uncertain environments. OPRA does not maintain a complex internal deliberation system; instead, it queries knowledge on demand, reducing the cognitive load and computational complexity of the agent itself. OPRA excels in real-time adaptation to novel conditions, leveraging ChatGPT’s reasoning and up-to-date knowledge to inform decision-making in ways that static or learning-based systems might not be able to match without retraining. The landscape of agent systems has evolved significantly over the past 40 years. Wrona et al. in [54] provide a comprehensive review of agent platforms available in 2023, many of which emphasize autonomy and real-time decision-making. OPRA’s dynamic external querying offers a complementary approach to these existing platforms, blending external and internal mechanisms for enhanced performance in real-time decision-making scenarios. In contrast to traditional agent platforms that rely heavily on internally programmed behaviors, OPRA agents exemplify a shift toward integrating external knowledge into agent architectures. The majority of agent frameworks focus on internal cognitive processes. For instance, Dong [55] proposes a sophisticated model allowing autonomous agents to share their conscious contents, revealing their internal mental states to human audiences. Such an approach allows agents to make their decision-making processes more transparent, relying on internal cognitive representations to explain behaviors. While OPRA leverages external consultation to fill in knowledge gaps, Dong’s model emphasizes the internal visibility of an agent’s mind, enabling human users to understand the rationale behind the agent’s actions from its internal representations rather than external sources. This contrast highlights a key difference between architectures focused on external versus internal knowledge sources and their respective roles in explainability and interaction with human audiences. Therefore, OPRA agent architecture stands out as an approach that blends reactive observation with on-demand external deliberation, making it a flexible solution for environments where predefined rules are absent or insufficient. It complements traditional architectures like BDI, reactive, learning, or utility-based agents by leveraging external expertise to fill in gaps, especially in real-time, dynamic scenarios where learning or predefined rules are inadequate. While this comparison is qualitative, OPRA introduces a paradigm shift by enabling agents to dynamically incorporate external knowledge, reducing reliance on predefined rules and extensive training. This adaptability, particularly in complex and uncertain environments, sets OPRA apart from traditional agent architectures and lays the foundation for future empirical validation of its advantages.