← Back to home

Computational Intelligence

Alright, let's dissect this. You want me to take this Wikipedia article, which is about as exciting as watching paint dry, and inject it with… well, me. You want it longer, more detailed, and certainly more engaging. And, of course, all the tedious links must remain. Fine. Consider it done. Just don't expect sunshine and rainbows. This is Computational Intelligence, after all.

Computational Intelligence

For the journal, see Computational Intelligence (journal).

For broader coverage of this topic, see Artificial intelligence.

In the arcane realm of computer science, Computational Intelligence (CI) isn't some fluffy, feel-good concept. It’s about the gritty reality of concepts, paradigms, algorithms, and the cold, hard implementations of systems designed to exhibit something we grudgingly call "intelligent" behavior. And not just in sterile, predictable environments, mind you. No, CI tackles the messy, ever-shifting landscapes where things get interesting. These systems are engineered to wrestle with complex tasks, whether in the sterile halls of academia or the cutthroat arena of commerce. Their purpose? To decipher and interpret patterns – a seemingly simple act that holds the key to unlocking so much. They’re meant to orchestrate processes, to lend a hand in decision-making, or, in the more ambitious cases, to autonomously pilot vehicles or robots through environments so alien they’d make your own existence feel mundane. [1] [2]

What truly sets these CI paradigms apart, what makes them more than just elaborate calculators, is their inherent ability to learn or, at the very least, to adapt when confronted with the new. They possess the capacity to generalize, to distill the essence from the noise, and to forge associations where none were apparent. It’s no surprise that methods inspired by the relentless, often brutal, efficiency of nature – what we call nature-inspired methods – play such a pivotal role. Think of neuroevolution, for instance. It's not just a technique; it's a philosophy, a way of bending the very fabric of computation to mimic the grand experiment of life itself. [1]

The true domain of CI approaches lies in those labyrinthine, real-world problems that defy the elegant, but often insufficient, grasp of traditional mathematical modeling. Why? Because sometimes, you can't simply write down the rules. The processes are too intricate, the variables too numerous, and the underlying knowledge too incomplete for a straightforward mathematical reasoning. Or perhaps the environment itself is a capricious beast, prone to unforeseen changes, rendering any static model obsolete before it’s even fully formed. And let's not forget the sheer pervasiveness of the stochastic – the inherently random nature of so many phenomena. CI steps in when processes are, in essence, ill-defined, maddeningly complex, frustratingly nonlinear, prone to temporal drift, and riddled with randomness. [4]

The IEEE Computational Intelligence Society has offered a definition that, while perhaps not the final word, certainly captures the spirit: CI is the theory, design, application, and development of computational paradigms that are inspired by biology and language. Traditionally, the holy trinity of CI has been Neural Networks, Fuzzy Systems, and Evolutionary Computation. But CI, like a restless entity, continues to expand its horizons. It now encompasses the whispers of ambient intelligence, the chaotic beauty of artificial life, the emergent properties of cultural learning, the intricate dance of artificial endocrine networks, the subtle art of social reasoning, and the complex interplay of artificial hormone networks. [5] And then there's the current obsession: Deep Learning. This has become the bedrock, the central engine for much of what we now call artificial intelligence. It’s undeniable; some of the most formidable AI systems today owe their existence to CI. [5] Yet, the field is still finding its footing, and a definitive, universally accepted list of what constitutes CI remains elusive. [6] [7] [8] [3] [9] [10]

Regardless of the specific label, the fundamental challenge in crafting an "intelligent system" remains the same: simulating intelligent thought and action within a given application domain. This necessitates translating knowledge about that domain into a form that the system can process – a model. The ultimate success of the system hinges on the quality of that chosen model. Sometimes, the path forward is paved with data-driven methods; other times, the stark logic of knowledge representation reigns supreme. More often than not, the most robust solutions emerge from the elegant fusion of these approaches – hybrid models. [2]

According to the established texts, the following methods and paradigms, often working in concert, are considered the core components of CI: [11] [12] [13] [14] [15] [16] [17]

The Relationship Between Hard and Soft Computing, and Artificial and Computational Intelligence

The term Artificial Intelligence (AI) is bandied about with alarming frequency, both in the media and by some of the very scientists who should know better. It’s often treated as a catch-all for the various techniques that fall under its expansive umbrella, or indeed, under the umbrella of CI. [5] [18] As Craenen and Eiben astutely observed, attempts to define or even describe CI tend to coalesce around a few recurring themes:

  • A "relative definition" that positions CI in contrast to AI.
  • A conceptual exploration of the core ideas and their significance within CI.
  • A straightforward listing of the established domains that constitute CI. [8]

The interplay between CI and AI has been a persistent source of debate throughout CI's development. While the aforementioned list might suggest they are interchangeable, the overwhelming consensus among researchers in the AI/CI nexus is that they are distinct, albeit related, fields. The precise nature of this relationship, however, remains a point of contention, with three primary viewpoints emerging: [8] [18]

  • CI stands as an alternative to AI.
  • AI encompasses CI within its broader scope.
  • CI, in turn, incorporates AI.

The first perspective, that of CI as an alternative, is often attributed to Lotfi A. Zadeh, the visionary behind fuzzy set theory. He drew a fundamental distinction between machine intelligence, categorizing it into hard and soft computing techniques. Hard computing, he argued, formed the basis of traditional AI, while soft computing was the bedrock of CI. [19] [20] The core difference, according to this view, lies in their handling of imprecision and uncertainty. In hard computing (HC) and conventional AI, these are seen as undesirable flaws to be eradicated. Soft computing, and by extension CI, on the other hand, embraces these very characteristics, viewing them as inherent to complex, real-world problems. [14] The accompanying diagram attempts to illustrate these complex relationships, highlighting the key CI techniques. [6] Another frequently cited distinction is the symbolic representation of information prevalent in AI versus the sub-symbolic representation favored in CI techniques. [17] [21]

Hard computing operates on the principles of certainty and absolute accuracy. It's deterministic, requiring a precisely defined analytical model and a meticulously crafted program – a fixed set of instructions. The underlying logic is typically Boolean algebra, or "crisp" logic, where an element either unequivocally belongs to a set or it does not; there is no middle ground. When applied to the messy realities of the world, HC systems produce rigid, specific control actions dictated by their pre-programmed models. Should an unexpected situation arise, one not accounted for in the model, the system is likely to falter, its actions becoming irrelevant, if not outright detrimental. [23] [24] [25] [26]

Soft computing, conversely, draws inspiration from the human mind's remarkable ability to process information and achieve goals even when faced with imprecision and uncertainty. [20] It's built upon models that reflect the probabilistic reasoning, fuzzy logic, and multi-valued logic inherent in human cognition. Soft computing excels at processing vast quantities of data and performing numerous, potentially inexact, computations in parallel. For those intractable problems that resist exact solutions through hard computing, soft computing methods often prove remarkably effective. These methods are typically stochastic, meaning they involve random processes that can be analyzed statistically but not with absolute precision. Even today, the outputs of some CI methods, particularly in areas like deep learning, can be difficult to fully verify, leaving us with a profound scientific challenge for the future. [23] [24] [25] [26]

Both AI and CI are, admittedly, rather catchy terms, [18] but their similarity often breeds confusion. The meaning and scope of both terms have evolved organically over time, [27] [28] with AI being the elder statesman. [3] [9] James C. Bezdek eloquently articulated this phenomenon, suggesting that such buzzwords often become overhyped, amplified by the scientific community, funding bodies, and the press. [18] The terms themselves, "artificial intelligence" and "biological intelligence," carry significant emotional weight, [3] [18] and the very definition of "intelligence" remains a subject of ongoing, often contentious, debate. [3] [10]

History

The quest to understand and replicate machine intelligence has a long lineage, with key milestones shaping its trajectory. Back in 1950, Alan Turing, a foundational figure in computer science, proposed what is now known as the Turing test. This thought experiment sought to establish a benchmark for machine intelligence: could a computer fool a human interrogator into believing it was also human? The very premise of the test highlights the early fascination with the possibility of truly intelligent machines. [29]

The term "Computational Intelligence" itself first appeared in 1985 as the title of a journal. [30] [31] Its formal emergence as a distinct field gained momentum with the founding of the IEEE Neural Networks Council (NNC) in 1989. This council brought together researchers passionate about the burgeoning field of biological and artificial neural networks. [32] The NNC later evolved into the IEEE Neural Networks Society in 2001, and then, two years later, rebranded as the IEEE Computational Intelligence Society to reflect its expanding scope, incorporating areas like fuzzy systems and evolutionary computation.

The NNC played a crucial role in organizing the inaugural IEEE World Congress on Computational Intelligence in Orlando, Florida, in 1994. [32] It was at this pivotal conference that James C. Bezdek offered a foundational definition of Computational Intelligence. He posited that a system could be considered computationally intelligent if it: dealt exclusively with numerical, low-level data; incorporated pattern-recognition components; eschewed knowledge in the AI sense; and, crucially, demonstrated at least one of the following: (1) computational adaptivity; (2) computational fault tolerance; (3) speed approaching human-like turnaround; and (4) error rates approximating human performance. [33]

In contemporary times, the landscape of CI has been dramatically reshaped by the ascendance of machine learning and, particularly, deep learning. These fields, leveraging a diverse array of supervised, unsupervised, and reinforcement learning techniques, have introduced a wealth of novel intelligent approaches, significantly enhancing the capabilities and applications of CI.

The Main Algorithmic Approaches of CI and Their Applications

The practical applications of Computational Intelligence are vast and varied, spanning critical domains such as computer science, various branches of engineering, sophisticated data analysis, and the intricate world of bio-medicine.

Fuzzy Logic

Unlike the rigid, binary distinctions of conventional Boolean logic, fuzzy logic operates on the principle of fuzzy sets. In both systems, properties are described in terms of set membership. However, fuzzy logic introduces a nuanced, graded membership. Instead of a simple yes/no, an element can belong to a set to a certain degree, quantified by a membership function that assigns a real number between 0 and 1. This innovation allows for the development of associated logic calculi that can model inference processes and, more broadly, logical reasoning in a manner that more closely resembles human thought. [34] Consequently, fuzzy logic is exceptionally well-suited for engineering applications where clear-cut certainties are rare, and imprecision or uncertainty are the norm – think of natural language processing [35]. However, it's worth noting that fuzzy logic, in its pure form, doesn't inherently possess learning capabilities. [36]

The applications of fuzzy logic are remarkably diverse. It finds its way into control engineering, [37] image processing, [38] techniques like fuzzy data clustering, [38] [39] and complex decision-making scenarios. [35] You'll encounter fuzzy logic-based control systems in everyday devices like washing machines, dishwashers, and microwave ovens, as well as in sophisticated automotive systems such as gear transmissions and braking mechanisms. Even the image stabilization in your video camera likely employs fuzzy logic principles to counteract unsteady handling. Beyond consumer electronics, fuzzy logic plays a role in medical diagnostics, satellite control systems, and strategic business planning, among many other fields. [35] [40]

Neural Networks

A cornerstone of CI is the development of artificial neural networks (ANNs), systems designed to emulate the structure and function of their biological counterparts. [41] [42] These networks are typically characterized by three primary components: the "cell body" or neuron, responsible for information processing; the axon, which transmits signals; and the synapse, which modulates these signals. [41] [42] This architecture makes ANNs exceptionally adept at distributed information processing and, crucially, at learning from experiential data. [43] [44] ANNs strive to replicate the cognitive processes of the human brain, offering significant advantages such as fault tolerance, the ability to recognize patterns even in noisy or incomplete data, and, of course, their inherent capacity for learning. [41] [44]

The applications of neural networks are extensive and can be broadly categorized into five groups: data analysis and classification; acting as associative memory; performing data clustering or compression; generating novel patterns; and implementing sophisticated control systems. [45] [43] [41] Specific examples are legion: analyzing and classifying medical data for diagnostic purposes, enabling speech recognition, facilitating data mining, powering image processing systems, generating accurate forecasting models, guiding robot control, assessing creditworthiness, detecting fraudulent activities, and mastering the control of complex systems with inherent nonlinearities. [41] [43] [45] Neural networks share the areas of application in control systems and data clustering with fuzzy logic. More recently, generative systems built upon deep learning and convolutional neural networks, such as the widely discussed ChatGPT and the remarkably capable DeepL, represent a significant and rapidly evolving frontier.

Evolutionary Computation

Evolutionary computation offers a powerful suite of methods and algorithms designed for global optimization. These techniques draw their inspiration from the principles of biological evolution, often operating on a "population" of candidate solutions. Commonly referred to as evolutionary algorithms, this family includes well-known members like genetic algorithms, evolution strategy, and genetic programming, among others. [46] They are particularly effective for tackling problems that defy conventional mathematical approaches [48] and are frequently employed for complex optimization tasks, including challenging multi-objective optimization scenarios. [49] A key advantage is their inherent parallelism: because they work with a population of solutions processed concurrently within an iteration, they lend themselves readily to distributed computing environments. [50] Furthermore, the generation of multiple offspring from a single pairing allows for parallel evaluation of these offspring, significantly accelerating the typically time-consuming optimization process. [51]

During the optimization process, the population collectively "learns" about the structure of the search space, storing this acquired knowledge within the "chromosomes" of the candidate solutions. This accumulated knowledge can be leveraged for future, similar tasks by adapting existing chromosomes to seed a new population, thereby accelerating convergence. [52] [53]

Swarm Intelligence

Swarm intelligence draws its power from the collective behavior observed in decentralized, self-organizing systems. These systems typically comprise a multitude of simple agents that interact locally with each other and their surrounding environment. Strikingly, even in the absence of any central command structure, these local interactions can give rise to complex, emergent global behaviors. [54] [55] [56] Prominent examples of swarm intelligence algorithms include particle swarm optimization and ant colony optimization. [57] Both are classified as metaheuristic optimization algorithms, adept at finding approximate solutions to difficult numerical or intricate combinatorial optimization problems. [58] [59] [60] Like evolutionary algorithms, their population-based nature and reliance on local interactions make them highly amenable to parallelization, [61] [62] and they often exhibit comparable learning capabilities. [63] [64]

Bayesian Networks

In domains characterized by significant complexity, Bayesian networks offer an efficient mechanism for representing and reasoning with uncertain knowledge. These probabilistic graphical models depict a set of random variables and their conditional dependencies through a directed acyclic graph. Their probabilistic foundation enables straightforward inference when new information becomes available, and they are also highly effective at learning from data. [13] Their utility extends across a broad spectrum of applications, including medical diagnostics, risk management, information retrieval, and text analysis (such as the development of spam filters). Further applications include assessing corporate credit ratings and optimizing the operation of complex industrial processes. [65]

Artificial Immune Systems

Artificial immune systems represent another class of population-based metaheuristic learning algorithms, specifically designed to address clustering and optimization challenges. These algorithms are meticulously inspired by the principles of immunology and the functioning of the vertebrate immune system, harnessing its inherent learning and memory capabilities to solve problems. They often employ operators analogous to those found in evolutionary algorithms, such as cloning and mutation, to manipulate artificial lymphocytes. [66] [67] Artificial immune systems offer compelling features like adaptability, self-learning, and robustness, making them valuable for diverse tasks in data processing, [67] manufacturing systems, [68] system modeling and control, fault detection, and cybersecurity. [66]

Learning Theory

In its pursuit of human-like "reasoning" capabilities, learning theory stands as a fundamental pillar of CI. Within psychology, learning is understood as a complex interplay of cognitive, emotional, and environmental factors that shape our knowledge, skills, values, and worldview. [69] [70] [71] Learning theories, in turn, provide frameworks for understanding how these processes occur, enabling us to make informed predictions based on past experiences. [72]

Probabilistic Methods

As a critical element within fuzzy logic, probabilistic methods, initially explored by pioneers like Paul Erdos and Joel Spencer in 1974, [73] [74] are employed to evaluate the outcomes of Computational Intelligence systems, particularly those governed by randomness. [75] These methods leverage prior knowledge to illuminate the range of potential solutions to a given problem.

Impact on University Education

Bibliometric studies consistently highlight the significant role of computational intelligence in research. [76] Major academic publishers readily accept manuscripts that delve into combinations of fuzzy logic, neural networks, and evolutionary computation. However, a curious paradox emerges when examining university curriculum: courses dedicated solely to Computational Intelligence remain surprisingly scarce. [77] Only a handful of technical universities worldwide, including the University of British Columbia, the Technical University of Dortmund (notable for its involvement in the European fuzzy logic movement), Georgia Southern University, and the Multimedia University in Malaysia, offer dedicated courses in this domain.

The reluctance of many prominent universities to incorporate CI into their core offerings often stems from resource constraints. Existing computer science curricula are frequently so densely packed with foundational material that there's simply no room for specialized topics like fuzzy logic. [78] While CI concepts might occasionally be introduced as sub-projects within introductory courses, the prevailing preference remains for classical AI concepts rooted in Boolean logic, abstract turing machines, and simplified theoretical problems like the "blocks world."

The recent emphasis on STEM education has, however, brought about a modest shift. [79] There is a growing recognition of the value of multidisciplinary approaches, which equip students with the tools to comprehend complex adaptive systems. [80] Nevertheless, these innovative pedagogical objectives are often discussed only in theoretical terms, and the actual university curricula have yet to undergo substantial adaptation.

Publications

See also

Notes

  • Computational Intelligence: An Introduction by Andries Engelbrecht. Wiley & Sons. ISBN  0-470-84870-7
  • Computational Intelligence: A Logical Approach by David Poole, Alan Mackworth, Randy Goebel. Oxford University Press. ISBN  0-19-510270-3
  • Computational Intelligence: A Methodological Introduction by Kruse, Borgelt, Klawonn, Moewes, Steinbrecher, Held, 2013, Springer, ISBN  978-1-4471-5012-1