Rational Machines and Artificial Intelligence

By Tshilidzi Marwala

Intelligent machines are populating our social, economic and political spaces. These intelligent machines are powered by Artificial Intelligence technologies such as deep learning. They are used in decision making. One element of decision making is the issue of rationality. Regulations such as the General Data Protection Regulation (GDPR) require that decisions that are made by these intelligent machines are explainable. Rational Machines and Artificial Intelligence proposes that explainable decisions are good but the explanation must be rational to prevent these decisions from being challenged. Noted author Tshilidzi Marwala studies the concept of machine rationality and compares this to the rationality bounds prescribed by Nobel Laureate Herbert Simon and rationality bounds derived from the work of Nobel Laureates Richard Thaler and Daniel Kahneman. Rational Machines and Artificial Intelligence describes why machine rationality is flexibly bounded due to advances in technology. This effectively means that optimally designed machines are more rational than human beings. Readers will also learn whether machine rationality can be quantified and identify how this can be achieved. Furthermore, the author discusses whether machine rationality is subjective. Finally, the author examines whether a population of intelligent machines collectively make more rational decisions than individual machines. Examples in biomedical engineering, social sciences and the financial sectors are used to illustrate these concepts.

• Provides an introduction to the key questions and challenges surrounding Rational Machines, including, When do we rely on decisions made by intelligent machines? What do decisions made by intelligent machines mean? Are these decisions rational or fair? Can we quantify these decisions? and Is rationality subjective?
• Introduces for the first time the concept of rational opportunity costs and the concept of flexibly bounded rationality as a rationality of intelligent machines and the implications of these issues on the reliability of machine decisions
• Includes coverage of Rational Counterfactuals, group versus individual rationality, and rational markets
• Discusses the application of Moore’s Law and advancements in Artificial Intelligence, as well as developments in the area of data acquisition and analysis technologies and how they affect the boundaries of intelligent machine rationality

• Language: English
• Publisher: Academic Press
• Release date: Mar 31, 2021
• ISBN: 9780128209448

Tshilidzi Marwala

Dr. Tshilidzi Marwala is the Rector of the United Nations (UN) University and the UN Under-Secretary-General from 1 March 2023. He was previously the Vice-Chancellor and Principal of the University of Johannesburg, Deputy Vice-Chancellor for Research and Executive Dean of the Faculty of Engineering at the University of Johannesburg. He was Associate Professor, Full Professor, the Carl and Emily Fuchs Chair of Systems and Control Engineering at the University of the Witwatersrand. He holds a Bachelor of Science in Mechanical Engineering (magna cum laude) from Case Western Reserve University, a Master of Mechanical Engineering from the University of Pretoria, PhD in Artificial Intelligence from the University of Cambridge and a Post-Doc at Imperial College (London). He is a registered professional engineer, a Fellow of TWAS (The World Academy of Sciences), the Academy of Science of South Africa, the African Academy of Sciences and the South African Academy of Engineering. He is a Senior Member of the IEEE and a distinguished member of the ACM. His research interests are multi-disciplinary and they include the theory and application of artificial intelligence to engineering, computer science, finance, social science and medicine. He has supervised 28 Doctoral students published 15 books in artificial intelligence (one translated into Chinese), over 300 papers in journals, proceedings, book chapters and magazines and holds five patents. He is an associate editor of the International Journal of Systems Science (Taylor and Francis Publishers). He has been a visiting scholar at Harvard University, University of California at Berkeley, Wolfson College of the University of Cambridge, Nanjing Tech University and Silesian University of Technology in Poland. His opinions have appeared in the New Scientist, The Economist, Time Magazine, BBC, CNN and the Oxford Union. Dr. Marwala is the author of Rational Machines and Artificial Intelligence from Elsevier Academic Press.

Book preview

Image 1

Preface

Tshilidzi Marwala, Department of Electrical Engineering, University of Johannesburg, Johannesburg, South Africa

Rationality is a concept that has been studied for more than 2000 years from the times of Socrates to the modern times of Daniel Kahneman. Rationality, simply put, means logical. Throughout these times, rationality has always been about humans. Now that intelligent machines that are powered by AI are becoming common, the rationality of machines is emerging as an essential area of study. Rational agents maximize their net utility. In this book, we use the concept of net utility, which is the difference between the expected utility and the utility cost needed to achieve that expected utility. The utility is the good that is derived from a particular action or object. According to Socrates, that which is good helps and preserves whereas that which is bad or evil corrupts and destroys. Accordingly, rationality is the maximization of that which helps and preserves. Rational or logical agents must meet two criteria: (1) they must achieve their goal or objective, which is to maximize the net utility and (2) they must use the minimum amount of energy to maximize the net utility. For an agent not to maximize the net utility, is like a businessman who leaves money at the marketplace because he has gathered enough for the day. Such a businessman is not acting logically or rationally. Furthermore, not using the minimum amount of energy is akin to going from Cape Town to Johannesburg via London for no other reason except that such a traveler wants to travel from Cape Town to Johannesburg.

This book studies the rationality of machines, especially in the light of the advances in AI, which are enabled by technologies such as deep learning, the abundance of data, and the increase in computational power due to Moore’s Law. This book studies the concept of bounded or limited rationality, which was proposed by the Nobel Laureate Herbert Simon, of both humans and machines and finds that the bounds or limits of rationality in machines are more flexible than those of humans. It also studies the theory of rational expectation of humans vs machines. The rational expectation is a theory that states that agents cannot be consistently wrong when predicting the future because it can always expand on the information it uses to predict the future. This book finds that machine rational expectation is more accurate than human rational expectation.

The book also studies the theory of rational choice. Rational choice states that if two options are presented and option A offers a higher expected net utility than option B, then a rational agent prefers option A over B. The situation becomes complex if viable options are numerous and unknown. Then when humans make such decisions, they also have to identify all alternative options, which can also be viewed as espousing opportunity costs, and this is computationally expensive. The process of imagining alternatives is called counterfactual thinking. When an intelligent machine is used, the capability to explore alternative options is enhanced and accelerated. Machines are found to be better rational choice agents than human beings. Furthermore, the book introduces rational counterfactuals which are counterfactuals that maximize the attainment of a particular goal. If the alternative option is as good as the most rational option, then this is a rational opportunity cost. The difference between rational opportunity costs and rational counterfactuals is that rational counterfactuals are imagined events and do not have to be practical while rational opportunity costs are viable and practical options that are forgone on making a choice. Machines were found to create more viable rational counterfactuals and opportunity costs than human beings.

Furthermore, this book studies the various properties of rationality. Can rationality be measured? Is rationality subjective? Is group rationality better than individual rationality? Are humans more rational than machines? Do machines make markets more rational than humans? Are machines more ethical than humans? This book concludes that it is difficult to measure rationality and therefore determines that rationality is subjective. Furthermore, it concludes that groups are more rational than individuals. It found that markets that are populated by machines are more rational and efficient than human-based markets. In conclusion, machines were found to be more rational than humans.

Chapter 1: Introduction to machine and human rationality

Abstract

This chapter presents the idea of machine rationality. It evaluates how intelligent machines are transforming the world. Some of the technologies that are transforming the world include artificial intelligence (AI), the Internet of things (IoT), robotics, advanced materials, and biotechnology. This transformation is collectively called the fourth industrial revolution (4IR). The 4IR has resulted in the emergence of intelligent machines of which machine rationality is an important characteristic. This chapter outlines the concept of machine rationality and expands on the first, second, and third industrial revolutions which as essential to understanding the 4IR.

Keywords

Rationality; Artificial intelligence; Intelligent machines; Fourth industrial revolution

1.1: Introduction

If we dig down to the linguistic definition of rationality, it is the use of information and logic to efficiently arrive at a conclusion. This chapter introduces the concept of machine rationality. It studies how intelligent machines are revolutionizing the world of politics, economics, and society. It studies some of the technologies that are driving intelligent machines such as artificial intelligence (AI), advanced materials, and biotechnology. These concepts are sometimes grouped and are collectively called the fourth industrial revolution (4IR) (Klein, 2008; Schwab, 2017). It studies the evolution of production, including the first, second, and the third industrial revolutions. It then links these to the emergence of intelligent machines and investigates an essential characteristic of an intelligent machine, which is machine rationality.

Perhaps the best manner to introduce the notion of rationality is to trace the fall of the Tsars. The Tsars ruled Russia with an iron fist from the 16th century until the communist takeover in 1917 (Ferro, 1995; Warnes, 1999). The word Tsar is derived from the word Caesar, the name of the famous Roman emperor Julius Caesar (Freeman, 2008). Tsar Nicholas II was the last Emperor of Russia and was the grandson of Queen Victoria of England. He married German princess Alexandra of Hesse who was also a granddaughter of Queen Victoria. Nicholas II and Alexandra were second cousins, making this union genetically irrational. In genetics, it is known that offspring out of blood relations can be dangerous as it can exacerbate family-related illnesses as it reduces genetic diversity.

A shift in a political regime is often explained through the concept of dialectical materialism. Dialectical materialism is the theoretical basis of communism, and it leads to the material interpretation of history, which, according to Karl Marx, is the history of class struggle (Mandel, 1977). Simply put, dialectical materialism infers that the type of society we live in is based on the mode of production that society adopts. This type of society continues until the internal contradictions, or in this case class struggles, within it create the material conditions for a revolution in order to change the identity of that society. Yet, the communists took over Russia, not because this concept was appealing to Russians, instead, Tsar Nicholas II was dethroned because of the failure of his reign due to him making too many mistakes, including marrying his cousin and fighting a disastrous war with Japan. Nicholas II failed because his life was full of irrationalities. I define irrationality, in this context as the act of contradicting the natural flow of things. In physics, there is a notion called the path of least resistance, which governs the dynamics of all things (Weissman, 2012) and explains why a given path is chosen in lieu of others. It is because of this concept of the least resistance that we have gravity. For instance, if you drop the ball, it falls along the shortest path to the ground under the force of gravity. It is also because of this concept that we have the principle of conservation of energy, which states that energy cannot be created or destroyed but can only be transferred from one form to another (Planck, 1923). Nicholas II’s life was so full of these contradictions that the communists exploited them before they overthrew him and killed his entire nuclear family.

But when exactly did things go wrong for Nicholas II? Firstly, by marrying his second cousin, the couple’s only son, Alexei—the heir to the Russian throne—had a family-inherited disease, hemophilia. Hemophilia is an inherited genetic illness in which the blood is unable to form clots and, therefore, prevent bleeding (Peyvandi et al., 2016). To treat this disease, the Romanovs (the Russian royal family) enlisted the services of a controversial faith healer Grigori Rasputin (Fuhrmann, 1990). Rasputin was not a doctor nor was he a scientist, and there is no evidence that indicates he learned to read and write. The Romanovs made another irrational decision, which was to entrust the care of their sick son to an illiterate spiritual man. Rasputin was a powermonger who became their advisor on matters of the state, including the military, much to the dismay of Prime Minister Alexander Trepov and the Commander-in-Chief Grand Duke Nicholas. Rasputin was so powerful that in the very sensitive era of the World War I, he effectively appointed Boris Sturmer as Prime Minister, and Minister of Interior and Minister of Foreign Affairs. Rasputin’s desire to maximize the control of Russia (which was rational from his perspective) was irrational for Tsar Nicholas II (because it weakened Russia and consequently his own rule). Because of the confluence of the irrationality of the Tsar when dealing with the Russian people, disease, wars, etc., the Romanovs lost their empire and ultimately their lives as the Red October swept through Russia (Ascher, 2014).

What could Tsar Nicholas II have done to prevent the disaster that befell his family? We call the hypothetical thinking of what could have been done a counterfactual (Rescher, 1964). If the identified counterfactual could have maximized Nicholas II’s stay in power, then this is a rational counterfactual (Marwala, 2014a). Counterfactuals are just thought experiments. The counterfactuals are real choices Nicholas II could have chosen, and forgoing such counterfactuals is called the opportunity cost. The opportunity cost that could have maximized the survival of the Tsar’s regime is the rational opportunity cost (Marwala and Hurwitz, 2017).

One type of counterfactual, which is the subject of this chapter, is a machine counterfactual. Suppose Tsar Nicholas II had an artificial intelligence (AI) machine to assist him in making all the decisions he had to make in his reign. What would the outcome of this have looked like? Would the communists have taken over Russia? Would Russia have industrialized as it did during Joseph Stalin’s reign? Would the World War II have been won? One clear thing is that the AI machine would not have relied on the conman, Rasputin. At the core of this AI machine is its rationality, which is the subject of the next section.

1.2: Machine rationality

What is machine rationality? Rational decision-making means making logical decisions and is the ideal concept of intelligence. In its purest form, rational agents maximize their utility. If one wanted to fly from Johannesburg to London, the choice that one makes would be rational if it maximizes comfort, minimizes costs, etc. Studies by Kahneman and Tversky have shown that human beings are at best irrational (Tversky and Kahneman, 1989). The reason why Tsar Nicholas II made so many irrational decisions was that he was a human being. Had he been an AI machine, the decisions that he made could have at least been more rational and perhaps because of this reason, his life could have been spared. A rational machine, i.e., AI, does tasks that an irrational machine, i.e., a human, cannot do. Rational decision means making logical decisions using information. Essentially, a rational agent processes all relevant information optimally to achieve its objective. Rationality has two fundamentals: the use of relevant information and the efficient processing of such information. Optimizing means a rational agent can find the minimum or maximum solutions. It means that such an agent, when traveling, can find the shortest distance between two locations, given the constraints. It means that when executing a task, it can identify where there would be minimum costs. Of course, there are complex mathematical arguments that will be explored in this book such as the condition in which it is possible to identify a globally optimum solution. Similarly, the efficient use of relevant and complete information is essential for an agent to act rationally. Complete relevant information is not practically attainable. The extent of the relevance and completeness of the information the AI machine uses as well as the efficiency of the intelligent machine in executing its task determine the quantification of the rationality of a machine. Unbounded rational decision-making is the notion of making decisions with perfect information, using a perfect brain in an optimized manner. Rational decision-making involves the optimal processing of complete information to achieve an objective. Because this entire process is optimized, it leads to the maximization of utility. Full or unbounded rationality is often unattainable and, consequently, rationality even of AI machines is always limited. Nobel Prize winner Herbert Simon called this phenomenon the theory of bounded rationality, which will be discussed later in this book (Simon, 1991).

1.3: Fourth industrial revolution

The 4IR is the era in which intelligent machines will perform tasks that were traditionally reserved for human beings (Marwala, 2020; Doorsamy et al., 2020). In the 4IR era, we should understand a human-machine as a system whose psychology, rationality, and effectiveness are interlinked. What then is the 4IR? For us to understand the 4IR, we should understand the previous industrial revolutions. The 4IR is the fourth because there were first, second, and third industrial revolutions.

The first industrial revolution occurred in England during the 17th century. We do not know the reason why the first industrial revolution happened in England. Given the population sizes of India and China at the time, the first industrial revolution should have happened in these two countries rather than in England. One of the reasons why it perhaps happened in England was because of Reformation, which led to the scientific revolution (Kuhn, 1962; Cameron, 2012), which saw developments in mathematics, physics, astronomy, biology, and chemistry. The Reformation led to the separation of Europe into the Catholic south and the Protestant north. To this day, the Catholic south is still more impoverished than the Protestant north. This scientific revolution in Britain gave us scientific luminaries such as Isaac Newton and James Watt. The first industrial revolution gave us the steam engine, which revolutionized transportation as well as the means and mode of production. Before that, manufacturing was performed by hand by trained craftsmen. The first industrial revolution allowed for commodities to be manufactured in bulk in factories. The steam trains gave rise to railroads that carried massive amounts of goods far more quickly and efficiently than could be done on horseback, for example.

The second industrial revolution saw the introduction of electricity and mass production. This changed the scale and speed of manufacturing significantly. It also gave us the electric motor, which in turn gave us the assembly line. The assembly line led to the mass production of goods and services. The second industrial revolution happened in the United States with ideas from Britain on electromagnetism (Darrigol, 2000). Michael Faraday realized that if one locates an electric conductor next to a magnet and moves the conductor, then electricity is generated. The reverse of that is that if one puts electricity in an electric conductor located next to a magnet, it moves, and this is the basis of an electric motor. The principle underlying this is called electromagnetism and was theorized by James Clerk Maxwell. Maxwell’s theory of electromagnetism was not consistent with Newtonian mechanics, so a Dutch physicist Lorentz developed correction factors (Lorentz transformation) to make the two theories consistent. It took Einstein’s theory of relativity to unify the theories of Newton, Maxwell, and Lorentz.

The third industrial revolution happened because of the discovery of semiconductors (Amos and James, 1999). These materials conduct electricity under certain conditions. Because of this reason, they can be used as efficient switches, and hence are used in digital computing. Before the advent of digital computers, we used to communicate digitally (using ones and zeros) using a telegram that used analogue switches. From the semiconductor devices, Bardeen, Brattain, and Shockley invented a transistor which ultimately led to the integrated circuit that makes modern computing possible. The third industrial revolution gave birth to the electronic age. Digital technology has improved so rapidly that every 2 years, we were able to double the processing power of computers, and this phenomenon is called Moore’s law. At some stage, we will not be able to miniaturize the integrated circuit because of the quantum effects, and this will be the end of Moore’s law.

The 4IR is an era characterized by the confluence of cyber, physical, and biological systems. An illustration of this is shown in Fig. 1.1. The cyber technologies include AI, blockchain, and the Internet of things (IoT) enabled by 5G technology. The developments in material science have led to the development of a new material called graphene as well as robots that can perform complex tasks in a hostile environment. On the biological technology side, developments in gene editing enhanced by AI are challenging the very essence of who we are as humans. An illustration of the evolution of 4IR is shown in Fig. 1.2.

Fig. 1.1

Fig. 1.1 An illustration of the fourth industrial revolution.

Fig. 1.2

Fig. 1.2 An illustration of the evolution of industrial revolutions.

We had seen automation in the previous three revolutions, but these were on a mechanical level—taking over the labor intensive skills. The 4IR, however, is an entire paradigm shift. The primary driver of the 4IR is AI, and this is the subject of the next section.

1.4: Artificial intelligence

AI is a computational technique that makes machines intelligent. While computers traditionally relied on people to tell them what to do and how to react, AI is based on machines that can learn and make their own decision. A machine is considered to be intelligent if it can analyze information and extract insights beyond the obvious. While computers traditionally relied on people to tell them what to do and how to react, AI is based on machines that can learn and make their own decision. The limit of how intelligent these machines can be is not known. However, we now know that AI machines can perform many complicated tasks such as playing better chess than human beings. There are three different types of AI: machine learning, soft computing, and computational intelligence (CI) (Bishop, 1995; Rutkowski, 2008; Chaturvedi, 2008). An illustration of AI with its subdisciplines is shown in Fig. 1.3. A type of machine learning is deep learning, which is a neural network with many layers. Machine learning is the use of data and statistics to create intelligent machines. Machine learning picks up on patterns and mimics human intelligence and in some instances, surpasses it. This gives AI some of the decision-making abilities that humans have. There are many types of machine learning, which include neural networks and support vector machines. The neural network was inspired by the structure of the human brain and consists of neurons that are connected, and they can map some inputs with outputs and is shown in Fig. 1.4.

Fig. 1.3

Fig. 1.3 Example of artificial intelligence.

Fig. 1.4

Fig. 1.4 A deep neural networks with three hidden layers.

For example, the input can be the face of a person, the output can be who that person is, and this is a face recognition algorithm.

CI uses group intelligence observed in nature to build intelligent machines. It is the technique of building intelligence by observing how nature works and using this information to make intelligent machines. Unlike neural network, which uses individual intelligence, CI uses group intelligence. An example of CI is ant colony optimization, which uses the principles observed in the workings of the ant colony to build intelligent machines. The first person to observe the intelligence and rationality of ants was a South African poet Eugene Marais who in his seminal book Die siel van die mier unlocked how white ants can build complicated anthills (Marais, 1937). These anthills have complicated tunnels with air conditioning systems, which are far better and efficient than the mechanical air conditioning systems in our rooms. An example of these ants and the corresponding ant hills are shown in Figs. 1.5 and 1.6. CI requires little data and sometimes no data at all. It has been used successfully for the clustering of machines (Xing et al., 2010a, b, c), designing cellular manufacturing layout (Xing et al., 2010a, b, c), cybersecurity (Ranjan et al., 2018), missing data estimation (Leke Betechouoh and Marwala, 2006), and designing storage and retrieval system (Xing et al., 2010a, b, c).

Fig. 1.5

Fig. 1.5 Ants that can build complex anthills.

Fig. 1.6

Fig. 1.6 Complex anthill built by ants.

Many other types of CI algorithms exploit group intelligence. One such CI algorithm, which uses the group intelligence of birds, is particle swarm optimization (PSO). This is based on the concept that individual solutions are particles that evolve into a more reliable solution. PSO is an intelligent global optimization algorithm that has been successfully used to tackle complex problems such as improving the accuracy of aircraft models (Mthembu et al., 2011; Boulkaibet et al., 2015) as well as predicting wind energy (Mbuvha et al., 2018).

Soft computing is an AI technique that requires limited data and is used to bring precision to issues such as linguistic variables. An example of soft computing is fuzzy logic. Fuzzy logic is based on vague, imprecise notions that may be true. Put differently, it is the logic of partial degrees of truth. In fuzzy logic, linguistic variables are encoded into the fuzzy domain using fuzzy membership functions, generating fuzzy rules and aggregating outputs and then defuzzification. This imitates the entire thought process a human being would have in a decision that would require a yes or a no answer. Another example of soft computing is a fuzzy-neuron system, which is an aggregation of fuzzy logic and neural networks. In other words, it uses a learning algorithm (neural network) to determine parameters (fuzzy logic) by processing data samples. The difference between fuzzy logic and neuro-fuzzy system is that the neuro-fuzzy system is more accurate than fuzzy logic, which is, in turn, more transparent, i.e., interpretable than the neuro-fuzzy system. The general law that governs fuzzy systems is that the more transparent it is, the less accurate it is whereas the less transparent it is, the more accurate it