Artificial Intelligence Theories: Application to CommonKAD Methodology

AI philosophy and relationship with design theory

Philosophies such as those proposed by Popper, Kuhn, and Lakatos apply broadly to most disciplines. In contrast, specific philosophies apply to disciplines like AI or related fields like computing and engineering. The AI philosophy, traditionally framed within the philosophy of science, should be expanded to include engineering and technology, contributing to the philosophy of engineering [18]. While transdisciplinarity is based on duality, [19] AI involves both duality and either–or logic. Design theory, central to the philosophy of technology [20], should be integrated into AI philosophy to encompass engineering aspects. Design theory, initially developed within design disciplines, has been adopted by fields such as AI as both science and engineering theory. Design science research frameworks contribute to creating new knowledge through scientific research methods and also innovative artifacts through engineering methods [21]. These design research methods are used in all computing disciplines [22], often being adapted or developed based on existing computing theories. Some of these are frameworks that, in addition to research, create artifact innovations. When theories and method-based theories are applied across computing disciplines, the underlying concepts are implicitly applied. Computing disciplines sometimes borrow methods from other disciplines and later adapt these methods to fit computing theories [22]. This ability to create theories and methods applicable to all computing disciplines shows that these disciplines have many similarities. This interconnectedness can be used by bioinspired design to create theories for one computing discipline from the existing theories of other computing disciplines.

Wieringa et al. [23] created the first SE design theory with inspiration from engineering theories, which are practical and widely usable as design theories. Similarly, AI design theories could be adapted to function as design research and innovation theories as demonstrated with IS design research [24]. This adaptation could streamline AI artifact research, development, and innovation.

The integration of philosophy and AI theory has become mainstream [25], with AI philosophy essential for neuroscience and neurophysiology to successfully reverse-engineer human and mammalian brains and to advance robotics [26]. Reverse engineering neuroscience and neurophysiology will benefit AI and other disciplines, especially through inspired design. AI theory is crucial for systematically transforming basic research into applied research and inventions, complementing AI philosophy.

The classical “computationalist” view is that AI and cognitive science are two sides of the same coin, with computing representations applicable across computing natural, artificial, and cognitive science systems [25], reflecting a transdisciplinary duality. Various paradigms, such as connectionist and symbolic, can design intelligent agents [27]. These paradigms may eventually be replaced by others such as behavioral robotics and neural networks [25]. Theories should ideally be derived from philosophy, and designers use design strategy to translate user and AI knowledge into AI products [28], leveraging multiple perspectives and alternatives.

Engineering design theories can be considered as user-friendly as they elucidate phenomena in terms of mechanisms, while most SE theories are statistical theories that are difficult to use [23]. The theories analyzed in this study are categorized as incomplete design theories to varying degrees. Previously, the author Wang [29] suggested that there are no widely used design theories in AI. In the aforementioned work, one of the potential reasons for this is that design theories are more difficult to use than design methods as design theories are formulated at a higher abstraction level than design methods. Another reason is that creating computing design theories is a more recent endeavor compared to the longstanding use of design methods. Currently, only pioneering innovators are experimenting with and adopting these theories, often through a process of trial and error and experimentation [30]. The lack of a critical mass of early adopters makes it difficult for subsequent groups of adopters as these groups rely on the knowledge and experience of early adopters.

Design theories have four characteristics: they address design issues, are generative, accept the propositions and language of other design theories, and utilize propositional logic [31]. These theories create solutions and designs, making them observable and understandable to adopters. AI design theory could incorporate elements from other design theories, building on existing research. The transfield between AI and SE could provide a basis for these disciplines to use each other’s theories [12].

How are AI and SE theories created?

Mature scientific disciplines build and accumulate knowledge primarily through theory development. In contrast, SE has historically placed less emphasis on theory building than these disciplines, and AI even less so [32]. For example, the annual SEMAT workshop focuses on general theory development in SE, starting with identifying the SE kernel common to all software engineering [33]. An equivalent focus on theory development in AI through workshops, conferences, or journals could play a similar role.

Science, engineering, and computing follow the same Technology Innovation, Development, and Theory Creation (TIDTC) cycle [5], which makes it possible for AI to learn from SE theory building. For example, the SEMAT SE kernel theory prevents teams from being constrained by methods by continuously improving them [33]. AI theories could similarly enhance the CommonKAD methodology. While there has been some borrowing between SE and AI, there is undoubtedly room for improvement. Borrowing theories is widely recognized as beneficial, though opinions differ on which and how many theories to borrow. The AI and SE communities should engage more through collaboration and share knowledge during major development and research projects across the TIDTC cycle steps: folklore, knowledge codification, and theory development.

The Function–Behavior–Structure (FBS) engineering design theory, initially from engineering [34], has been borrowed by SE. Although there are few studies on theory borrowing in computing, IS have conducted several investigations. For example, studies have found that even inappropriate theory borrowing may be useful as a learning process, developing theory-borrowing skills [35]. The authors Hall and Rapanotti [36] have found that borrowing similar ideas and knowledge from IS by SE is useful, and similarities between computing disciplines can be explored for potential borrowing. SE’s rational unified process, Pahl engineering, and service design models when cast into FBS revealed many similarities, indicating that design may be independent of the discipline [37]. This disciplinary independence is a transdisciplinary principle. The similarities between design theories and some design elements across disciplines suggest opportunities for borrowing and extending some computing design theories. In the study by Ralph [34], combining SE sense-making with FBS design theories for modeling software systems has been proposed. The FBS theory proponents claim that the theory is based on eight fundamental design processes [34, 38]. In addition, FBS has been applied in different disciplines, providing some evidence supporting this claim.

Theories make knowledge more transferable [39], and are important for conceptualization and communication within a research field, aiding practitioners in making strategic technology and project decisions [40]. The SE theories have a higher degree of diversity than those of physics [39], but transdisciplinarity can help reduce this variation [12]. AI is transdisciplinary [41], with a broader range of theories compared to SE. The goal of intelligent SE is to apply SE knowledge to AI and vice versa [42], increasing opportunities for theory use in both disciplines. Figure 1 shows the evolution of theories and methods.

Furthermore, the author Gruner [43] discussed that general theories of SE and IS design could not exist, and any attempt to develop such theories is ineffective and wastes resources. Also, the authors Shaw and Garlan [5] discussed developing SE to mature the discipline into a true engineering field. Furthermore, in the study by Gruner [43], the development of microtheories has been supported. In contrast, Shaw and Garlan [5] advocate for three types of theories: general, mid-range, and microtheories. There is a contradiction between these two studies on general theories and agreement on microtheories. Moreover, Gruner [43] does not consider mid-range theories. The contradiction can be resolved with SE dialectics. In addition, Shaw and Garlan [5] present the thesis for general theories while Gruner presents the antithesis. Another study supporting the thesis for the development of general theories proposed a unified theory of intelligence, unifying artificial, human, and other types of intelligence [44]. A synthesis could resolve the contradiction by, for example, stating situations when developing general theories is appropriate and when it is not.

Premature theorizing is likely to be wrong but not fruitless and is preferable to delayed theorizing, which can result in more significant issues [32]. Failure and success in theorizing produces useful lessons and helps develop theory-building skills. The same applies to borrowing the wrong theories [35]. Failure helps discover what does not work while delayed theorizing prolongs the creation of knowledge on shaky foundations, leading to diminishing returns on research efforts.

Furthermore, Shaw and Garlan [5] developed their general science–engineering cycle theory method by using analogies from engineering. General theories of theory building apply across science, engineering, and technology disciplines and, therefore, to AI. Some non-general engineering and software engineering theories can be generalized or unified to make them applicable to AI.

Kernel artificial intelligence theories

The kernel theories are general and cover a broad range of phenomena. Three kernel AI theories are analyzed here.

Intelligent decision-making theory (IDMT) unifies decision-making theory with known probability distribution and Solomonoff’s universal induction theory with unknown probability distribution [45]. Unification enables the theory to deal with problems with known and unknown probability distributions. The IDMT is agent-oriented and based on a universal agent that learns by inferring and acting using decision theory and induction.

Unified intelligence theory (UIT) is a kernel theory unifying human, biological, and cognitive intelligence [44]. It is transdisciplinary since it transcends different intelligence disciplines. It unifies knowledge from different disciplines of computing, AI, psychology, and cognitive science. Transdisciplinary principles applied to create theory are unifying knowledge, dealing with socially relevant issues, and transcending disciplines [46]. The theory deals with socially relevant issues by using human intelligence.

The core of UIT theory is abstract intelligence, consisting of a mechanism of models and types of intelligence components and their interactions. The interactions are either direct or indirect through other components. Abstract intelligence provides mechanisms for advanced intelligence, for example, thinking, inference, and perception [44]. The UIT has functional, logical, cognitive, and neural models. Natural intelligence interacts with abstract intelligence through AI. Computational intelligence interacts with abstract intelligence through machinable intelligence. Also, UIT is a design theory that has a defined architecture and environment for intelligent systems carrying out transformations and implementing mechanisms of phenomena [23] in all its components. Humans and intelligent systems are environmental components that carry out transformation. However, the theory does not entirely explain how it is accomplished. Its architecture provides the structure and interaction between its components.

Mathematical intelligence structures (MIS) theory provides a mathematical framework for natural, biological, and artificial intelligence, aiming to collapse the distinction between human brains and artificial mechanisms [47]. Its mathematical formulation makes it easy to apply simulation. However, it primarily focuses on human intelligence, with limited integration of other biological or non-biological intelligence. It is not clear how non-biological intelligence can be integrated into the three kernel theories.

Task theories

Two task-based theories are principles task theory (PTT) [48] and classification task theory (CTT) [49]. These theories focus on dividing the activities required for developing a computing system into tasks, the smallest atomic units of work assignable to a team member. The PTT supports the abstraction and concretization of tasks, differentiating between them, determining complexity, estimating required resources, and constructing tasks [48]. It is based on three task mechanisms: casual relations, task difficulty, and level of detail [50]. Activities are divided into tasks, establishing relationships between those tasks, while principles are the basis of reasoning that guides how actions are executed to perform tasks.

Furthermore, CTT supports AI system development by classifying tasks based on their roles in constructing a system. The theory is based on generic operations required to construct a system by considering input, output, and process by asking what problems the system can solve in mechanical, electronic, and biological systems [51]. The classification is based on system theory: input, output, and process. The tasks are classified into two categories, interpret and construct, each of which is divided into subcategories [49]. Interpret involves identifying, predicting, and controlling while construct involves specifying, designing, and assembling.

The PTT is a partial design theory because it focuses primarily on mechanism and environmental transformational components. Also, CTT has mechanism and environmental transformational components, making it a partial design theory as well.

Innovation theory

Creativity is a fundamental feature of human intelligence that is challenging for AI to replicate [52]. Machines are still lagging behind human creativity, invention, and innovation in most cases. However, in some areas, AI creativity and innovation are useful in specific contexts. The synergy between AI and human creativity and innovation can be leveraged through human–AI cocreation systems.

AI innovation theory consists of three innovation theories: combinational, transformational, and exploratory [53]. Combinatorial innovation combines existing elements in new ways. Transformational innovation transforms concept space. Transformational innovation applies when concept space is insufficient for innovation [54], which requires adding, dropping, and modifying concepts and relationships to make the concept space sufficient. Exploratory innovation involves exploring the innovation landscape to discover new elements required for innovation.

The Boden theory was used to develop a creative problem-solving framework [54] for creating innovative artifacts. This framework involves problem formulation through planning and learning, with knowledge represented using both symbolic and non-symbolic formalisms. The concept of space is manipulated using various innovation theories. Researchers investigated how to integrate Boden’s innovation theories into this framework and ultimately created it to facilitate innovative thinking and problem-solving.

Artificial intelligence decision design theory

Crowdsourcing system design theory combines collective intelligence (CI) and AI to support decision-makers in evaluating large amounts of user-generated content [56]. Crowdsourcing is a way of getting ideas from a distributed system of people. Design theory is being used in AI and AI in design, and both to improve systems [57]. Design theory is for developing innovative products.

CI is an emergent property of a distributed system of people using their combined synergized intelligence. Motivation and trust are required by humans to apply CI to design intelligent systems like robots [58]. Most AI systems are based on human individual intelligence or derived intelligence of individuals or communities of organisms, like in ant colony algorithms. Human individuals of a CI system interacting with each other form an adaptive system [58] whose properties cannot be understood by what individuals do but rather by the total sum of their actions.

Genetic programming theory

Genetic theory is a genetics bioinspired design theory. It is based on schemata theory [59] and evolutionary algorithms [60] that subdivide the search space into subset spaces called schemata [59]. Schemata subspace programs are evolved using generic operators of selection, crossover, and mutation. Selection determines solutions to retain by eliminating a few of the worst solutions from the solution population. Crossover combines parts of existing solutions to create new solutions while mutation generates random novel solution parts. Genetic algorithms have been used to design and model markets, innovation systems, and other types of systems.

Ways and challenges of integrating theories and CommonKAD

One loose way of integrating theories and CommonKAD is applying theories at the beginning of each CommonKAD phase. Another way of loosely integrating them is letting developers decide when to apply theories during development processes. Loosely integrating theories increases the range of problems they can solve. Alternatively, tight integration has the advantage of increasing developer productivity and the disadvantages of reducing the number of problems the combination can solve and the creativity developers can leverage because tight combination leaves little room for flexibility.

CommonKAD is a heavyweight, plan-based methodology that emphasizes comprehensive planning, analysis, design, and documentation. Integrating theories with CommonKAD has both advantages and disadvantages. For example, it may increase overhead costs like project costs beyond what is allowed by small-to-medium AI systems’ project budgets [61]. One way to overcome this is by applying the modeling principle of traveling light by creating only minimum viable models that make it easier and faster to create software [62]. Another way is by downscaling CommonKAD to make it lightweight, reducing overheads in developing small-to-medium AI systems.

CommonKAD is modeling-oriented, whereas theories are not. Integrating theories and methodology would require orienting theories towards modeling to create a seamless development process, avoiding the need to switch between modeling and non-modeling orientations. Switching between orientations makes development difficult, requiring more effort and consequently increasing development costs.

Combining theories into a system of theories would create a more powerful way of understanding and improving CommonKAD methodology and other AI artifacts. There are useful approaches combining computing into a system of theories and methods that create a theory nexus [63] and a core design theory based on common field elements [33]. A CommonKAD theory nexus would allow adding more AI and SE methods to the nexus. SE agile methods are the most commonly combined methods in full or partial phases with AI methods for developing AI-based systems [64]. A theory nexus is a loose combination that enables the application of parts of one or more methods on task by leveraging their strengths and counteracting their weaknesses. This gives innovators more freedom to be creative. Nexus solves the problem of having to tackle tasks in fixed ways. Design nexus can be made more innovative by integrating it with AI innovation theories. Methods follow design theories [36, 65]. Among design considerations for creating a design nexus, design theories that methods can follow synergistically are included.