Introduction to AI
Along with the increasingly dynamic development of new technologies in the 21st century, the term artificial intelligence has become more and more common. While in the first decade, for many people, artificial intelligence was rather synonymous with a robot that looks, thinks, and acts like a human, today, there is a tendency to degenerate this concept accordingly. in which area it is used. It even gets to the point that many organizations use it for marketing purposes. For example, the automation of a process involving a computing system that calculates faster than an employee is called artificial intelligence. This is done in order to promote the company and its activities, because of the idea that it uses the so-called smart component.
To answer the question of what artificial intelligence really is, it is good to consider what it is not. Often the definitions of AI come from companies, electronic publications (articles, magazines, etc.) or books. In many of them, AI is seen as a branch of computer science or as a set of computer science techniques that allow developed systems to perform activities that typically require human intelligence. Artificial intelligence is found and present in many areas of scientific knowledge, respectively, it can be argued that AI is an interdisciplinary science. For this reason, it would be wrong to claim that it is a type of computer science. As another argument in support of this, it can be pointed out that if AI is considered as a system - a type of algorithm and understands its essence, it will be found that the algorithm is built on some logic and mathematical operations, things that come out outside the scope of computer science.
Research formulation
Within the present research goal is to focus on the narrow AI, as means to enhance human (and business) capabilities by assisting in achieving various low-level tasks.
This research goal includes solving the following research objectives:
- Studying the fundamental concepts of A.I. systems.
- Exhaustively describe the essence of self-organization and its methodological aspects.
- Compare self-organization and self-learning, and review the convergent techniques.
- Develop new self-organization methods.
- Implement self-organization solutions in modern software environments.
- Develop usecase solutions for:
- automation of financial and personnel decisions;
- predictive modeling of complex behaviors;
- self-perfecting systems for education.
Focusing on “narrow A.I.” and on automation of low-level operations does not mean that the objectives of the proposed research are narrow or low-level. On the contrary – on could easily make an argument that the impact may be global even by such standards as the UN Sustainable Development Goals (United Nations, 2018).
All A.I. research in general would work indirectly toward achieving majority of aforementioned goals (to name a few: 3. Healthcare and well-being at all ages, 4. Quality education and lifelong learning, 7. Affordable, reliable, and sustainable energy for all, 10. Reduce inequality among countries, 11. Sustainable cities, 12. Sustainable production).
The authors of the current proposal see the biggest direct impact in the domain of goal 8. Promote sustained, inclusive, and sustainable economic growth, full and productive employment, and decent work for all (specifically target 8.2. Achieve higher levels of economic productivity through diversification, technological upgrading, and innovation, including through a focus on high value added and labor-intensive sectors) and goal 9. Build resilient infrastructure, promote inclusive and sustainable industrialization and foster innovation (specifically target 9.5. Enhance scientific research, upgrade the technological capabilities of industrial sectors in all countries, in particular developing countries, including, by 2030, encouraging innovation and substantially increasing the number of research and development workers per 1 million people and public and private research and development spending; and target 9.b. Support domestic technology development, research, and innovation in developing countries, including by ensuring a conducive policy environment for, inter alia, industrial diversification and value addition to commodities).
Types of self-perfection A.I. systems
As an effort to keep up with the technological advancements as well as trying to keep up with ever-more emerging complexity, there is now a new tendency among researchers to propose autonomous solutions for low-end business tasks. The general approach preached by the authors is based on the idea of self-perfection, common to the field of cybernetics.
In order to prolong the successful business operations, the controlling system should be adaptable i.e., it should respond to the changes by dynamically changing its structure and / or the values of the significant variables. In other words , the non-stationarity of the environment may (and should) be compensated through self-perfecting of the controlling system. Self-perfection has three important possible features - Self-learning, Self- organization, and Automation.
Automation is the minimization of human intervention to the operations management process, which is done mostly by a computerized controlling system. Automation is the very essence behind cybernetics and control theory and it greatly aides the processes of self-learning and self-organization (namely the “self-” part).
Self-learning (also known as “machine learning”) is the process of adjusting the internal variables of the controlling system by some sort of algorithm or a procedure, so that the significant outcomes from the controlled system improve (approach the goals) while the system is functioning. The process is known as parametric identification of a system.
Self-organization is the process of rearranging and reformatting the internal structure (subsystems and connections among them) of the controlling system by some sort of algorithm or a procedure, so that the significant outcomes approach the goals (Marchev, Motzev, 1983). The process is known as structure and parametric identification of a system.
Comparison of selected features of Machine learning and Self-organization (Motzev, 2016)
The conceptual evolution of artificial intelligence has begun much earlier than the invention of computers with many attempts in art and engineering to simulate artificial behavior of various types. But even after the invention of advanced computing the roadmap is still windy and inconclusive. As the predominant doctrine nowadays is machine learning (also known as self-learning), it is by no means the ultimate universal solution.
As a main thesis, this research tries to prove the greater potential of self-organization over machine learning as a method to achieve truly autonomous decision-making methods.
Concepts of self-organization
In an attempt to introduce the readers with the ways to pursue the research objectives and to prove the main thesis the next paragraphs will showcase the fundamental concepts of self-organization in a multi-disciplinary overview.
The science of cybernetics is transdisciplinary and complex by nature. This makes it suitable for solving complex and dynamic problems. Norbert Wiener had noticed that the feedback principle is a key feature for every system, which can change their actions in response to the environment. He developed his concept into the field of cybernetics and defined it in 1948 as "the scientific study of control and communication in the animal and the machine." Feedback measures the performance of a system by constantly comparing actual and desired behavior and it is adjusting the control action (Wiener, 1948).
Cybernetics concentrates on the function and behavior of systems in general. It has the desire to describe and control the dynamic and probabilistic behavior of every complex system. Through cybernetics Norbert Wiener searched for a universal theory of knowledge, order, and calculation. His theories of feedback, information, and their statistical measurement have become essential components for engineering, robotics, and human sciences.
Most of the systems modeled by traditional mathematical methods are linear - all effects are proportional to their causes. In complex systems, the relation between cause and effect is not straightforward. Small cause could have a large effect. Each component in a system affects the other elements and to control and understand that the effect is used the feedback.
The control systems which are regulated by feedback loops (also known as cybernetic loop) are the primary object of study within cybernetics. Wiener was inspired by machines that could self-correct, modify, their own behavior in response to external stimulations. He defines feedback in general as “the chain of transmission and return of information.” The feedback loop is the basic unit of control systems.
- input, which provides the system with information about the environment;
- a reference value which indicates the goal;
- in case of a discrepancy between input and reference value - an output function that causes the system to ‘behave’ in a way that affects the environment in order to reduce the discrepancy.
The feedback loop minimizes the difference between the current situation and the goal of the controlled system. Feedback has two values - positive (amplifies the initial change) and negative (stabilizes the system by bringing deviations back to their original state). (Schneider, ME) (2015)
In self-organizing systems, the control function is distributed over the whole of the system. In more complex systems the controlling system needs a self-perfection block, which would focus on finding effective control methods according to predefined requirements and specific goals. Based on the complexity of the system there will be many interlocked positive and negative feedback loops - in some directions the change should be amplified, while in others it should be suppressed.
Regarding a cybernetic system, variety is a descriptor of the number of possible states within the system. Every cybernetic system varies widely in size but also in complexity - variety grows with the size of the systems and its processes and it is creating complexity. Complex systems must be able to deal with a greater variety of disturbance. In 1956 W. Ross Ashby stated that only “variety can destroy variety”.
In other words - the controlling system has to feature equal or higher variety than the controlled system, it must contain as much variety as the controlled system. The actions of the controller must have more effect on the state of the controlled system, not the other way around. Roger C. Conant and W. Ross Ashby formulated a theorem which stated: "Every good regulator of a system must be a model of that system" (1970).
This theorem can be applied to all regulating and self-regulation systems. Self-regulating system is a system where there is no external control, but local interactions between the system components which leads to global regularities. According to Bar-Yam (2004), an efficient controller requires at least the same complexity as the complexity of the controlled system. In order to face the emergence and self-organization of the controlled system, the controller requires a balance between predictability and adaptability. Gershenson , C. (2013).
The very essence of cybernetics is directed towards self-organization and adaptability through evolution. To be considered as a complex, the system requires both emergence and self-organization. And to be considered as a self-organizing, the system should increase its own organization. Self-organizing systems learn from their environments, evoke new interaction rules between the components and rearrange their internal structure in order to achieve a result in the desired direction.
In 1962 W. Ross Ashby suggested his conception of self-organization: “every isolated, determinate, dynamic system obeying unchanging laws will develop organisms that are adapted to their environments”. He noted that every dynamic system tends to evolve through the state of equilibrium. This reduces the uncertainty and entropy of the system.
Ashby distinguished two meanings of the term “ self-organizing system” . First, the system can be concerned as a self-organizing if it contains a separate, independent element which can join through certain rules and define different structure. A disordered system is a system where independent components act in random ways; they do not influence each other. Due to the controller (feedback) the system can change its internal configuration. Creating order, the elements of the system are starting to correlate. This correlation is a useful measure to study the system transition from one state to another, the transition “from unorganized to organized”.
Second, self-organization means to rearrange the system structure with a concrete goal. By the feedback loops, the system generates its own intelligence. It searches independently for better solutions. When a bad organized internal elements and procedures are ordering in a new more useful configuration which leads to the desired outcome - the system is "changing from a bad organization to a good one."
The advantage of self-organizing systems is that their autonomy allows independent search for solutions. The interaction with the environment makes the systems flexible and adaptive. The feedback loop allows them to evolve and adapt to a constantly changing environment, so they can react very quickly to every change.
A particular combination of self-organization and emergence gives rise to self-reproduction. If a system is capable of reproducing itself autonomously using raw elements, it is a self-reproducing system. It is important to distinguish both terms self-reproduction and self-replication. Self-replication means to create a very faithful copy, which requires very complex control and editing. Self-reproduction processes are less precise, it refers to a statistical process of making very similar things. . Studying artificial self-replicating systems began with the desire to understand deeper complex systems and the fundamentals of information-processing principles. (Von Neumann, 1961).
Self-organization allows a system to develop its structure autonomously. The selection of the elements is responsible for the system's behavior and adaptation to the changing environment. Even in 1948 John von Neumann formalized his idea of a cellular automata in order to create a model for self-reproducing machine and to understand how evolution works. The cellular automata are one of the best examples of how complex dynamics can emerge from a system of many simple interconnected components.
Unlike Turing's automata, where the outputs are zeros and ones Turing, AM (1936), the goal of Neumann is to create automation where the output is another automation - self-reproduced machine without loss of complexity. The center of his interest was not to reproduce intelligent behaviors, but a self-replicating machine. He was never able to build his self-reproducing machine, but his concept became widely used. Neumann's abstract model of self-reproducing machine is based on the cell reproduction concept. Each cell has a state, and each cell is connected to certain neighboring cells. The states of the cells change at discrete time-steps. By predefined rules regarding previous states and current neighbors the cellular automaton passes from one state to another. In “Theory of Self-Reproducing Automata” (1966), Neumann proved that the dynamics in a cellular automaton are similar to the biological processes of self-reproduction and evolution.
Although the idea was implemented differently, Alan Turing, John von Neumann and Norbert Wiener were the first to discuss the future machine than can be learned by itself. There are some major differences in their concepts. Turing wanted to create an intelligent machine that can imitate human tasks and behavior. Wiener wanted to achieve effective autonomous communication between machines. Von Neumann's goal was to create self-reproducing automation, which increases its complexity with each generation.
Alan Turing declares that what he wants is to create a machine that can learn from experience - to create a “ learning machine ” In 1950 he published the first major academic paper on the foundations of machine learning and artificial intelligence. He assumed if a machine could imitate a person there are three main components - initial state, education, and other experience.
He proposed the idea to use “child machines” as initial elements, which can learn from rewards and punishment. These machines would inherit their structure from their parents, the previous generation. During the self-reproduction, the changes in the structure would be caused by some sort of mutations.
Turing suggests that the population depends on natural selection and survival of the fittest. Once a new population is created after a mutation, its members must be evaluated according to a fitness function and only the top-scoring members of the offspring are chosen to continue. In the selection process intelligence will accumulate and each new generation would be better than the previous one.
In natural selection chances of surviving increased, but it is a slow and long process. The selection is based on the adaptable character who is able to cooperate with the environment. In order for faster evolution in a direction closer to the system's goal, it is better to use artificial selection (directed selection). This is a controlled selection which involves an artificial process where the selection is done by favoring the desired characters of the offspring. It is a set of rules that choose who will be the parents of the next generation. Every element of the offspring which cross the requirements is moved forward to reproduce and generate new generation. This helps to bring a new variety of characteristics in the system components. Artificial selection is done to enhance the quality of the generation, to achieve ultimate solutions.