3

Artificial Intelligence

Computer systems are installed at airports to sniff luggage forexplosives. Military hardware is becoming increasingly reliant onresearch into intelligent machinery: missiles now find their targets withthe aid of machine vision systems.

Over the past half-century there has been intense research into theconstruction of intelligent machinery – the problem of creating Artificial

Intelligence. This research has resulted in chess-playing computerscapable of beating the best players, and humanoid robots able tonegotiate novel environments and interact with people.

Many advances have practical applications …

Computer systems can extractknowledge from giganticcollections of data to helpscientists discover new drugtreatments.

Intelligent machinery ...

... can mean life or death.

4

Defining the AI Problem

Research into Artificial Intelligence, or AI, has resulted in successfulengineering projects. But perhaps more importantly, AI raises questionsthat extend way beyond engineering applications.

The capabilities of an agent could extend beyond that which we cancurrently imagine. This is an exceptionally bold enterprise which tackles,head-on, philosophical arguments which have been raging forthousands of years.

The holy grail of Artificial Intelligence isto understand man as a machine.

Artificial Intelligence also aims to arrive at a generaltheory of intelligent action in agents: not just humans andanimals, but individuals in the wider sense.

5

What Is an Agent

An agent is something capable of intelligent behaviour. It could be arobot or a computer program. Physical agents, such as robots, have aclear interpretation. They are realized as a physical device that interactswith a physical environment. The majority of AI research, however, isconcerned with virtual or software agents that exist as models occupyinga virtual environment held inside a computer.

Some AI systems solve problems by employing techniques observed inant colonies. So, in this case, what appears to be a single agent may berelying on the combined behaviour of hundreds of sub-agents.

The distinction between physical andvirtual agents is not always clear.

Researchers may experiment with virtualagents that occasionally becomephysically instantiated by downloadingthemselves into a robotic body.

An agent itself may also becomposed of many sub-agents.

6

AI as an Empirical Science

Artificial Intelligence is a huge undertaking. Marvin Minsky (b. 1927),one of the founding fathers of AI, argues: “The AI problem is one of thehardest science has ever undertaken.” AI has one foot in science andone in engineering.

In its most extreme form, known as Strong AI,the goal is to build a machine capable of thought,consciousness and emotions. This view holds thathumans are no more than elaborate computers.

Weak AI is lessaudacious.

7

The aim of Weak AI is to develop theories of human and animalintelligence, and then test these theories by building working models,usually in the form of computer programs or robots.

So, for Weak AI, the model is a useful tool for understanding the mind;for Strong AI, the model is a mind.

It is not proposed that machinesthemselves are capable of thought,consciousness and emotions.

The AI researcher viewsthe working model as atool to aid understanding.

8

Alien-AI Engineering

AI also aims to build machinery that is not necessarily based on humanor animal intelligence.

Because the mechanisms underlying such systems are not intended tomirror the mechanisms underlying human intelligence, this approach toAI is sometimes termed Alien-AI.

Such machines may exhibit intelligent behaviour, butthe basis for this behaviour is not important.

The aim is to design useful intelligentmachinery by whatever means.

9

Solving the AI Problem

So, for some, solving the AI problem would mean finding a way to buildmachines with capabilities on a par with, or beyond, those found inhumans.

But for most researchers working on AI, the outcome of the Strong AIdebate is of little direct consequence.

The goal of Strong AI is subject to heateddebate and may turn out to be impossible.

Humans and animals mayturn out to be the leastintelligent examples of aclass of intelligent agentsyet to be discovered.

10

Ambition Within Limits

AI, in its weak form, concerns itself more with the degree to which wecan explain the mechanisms that underlie human and animal behaviour.

The strong stance can be contrasted with the more widespread andcautious goal of engineering clever machines, which is already anestablished approach, proven by successful engineering projects.

The construction ofintelligent machinesis used as a vehiclefor understandingintelligent action.

Strong AI is highly ambitious and sets itselfgoals that may be beyond our grasp.

11

Taking AI to its Limits

Immortality and Transhumanism

“We cannot hold back AI any more than primitive man could have

suppressed the spread of speaking” – Doug Lenat and EdwardFeigenbaum

If we assume that Strong AI is a real possibility, then severalfundamental questions emerge.

The problem that Strong AI aims to solve must shed light on thispossibility. Strong AI’s hypothesis is that thought, as well as other mentalcharacteristics, is not inextricably linked to our organic bodies. Thismakes immortality a possibility, because one’s mental life could exist ona more robust platform.

Imagine being able to leave your body and shifting your mentallife onto machinery that has better long-term prospects than theconstantly ageing organic body you currently inhabit.

This possibility isentertained byTranshumanistsand Extropians.

12

Super-Human Intelligence

Perhaps our intellectual capacity is limited by the design of our brain.Our brain structure has evolved over millions of years. There isabsolutely no reason to presume it cannot evolve further, either throughcontinued biological evolution or as a result of human interventionthrough engineering. The job our brain does is amazing when weconsider that the machinery it is made from is very slow in comparisonto the cheap electrical components that make up a modern computer.

For some, this is oneof the goals of AI.

Brains built from moreadvanced machinery couldresult in “super-humanintelligence”.

13

Neighbouring Disciplines

“Certum quod factum.” [One is certain only of what one builds] –Giambattista Vico (1668–1744)

What sets AI apart from other attempts to understand the mechanismsbehind human and animal cognition is that AI aims to gainunderstanding by building working models. Through the syntheticconstruction of working models, AI can test and develop theories ofintelligent action.

AI’s goal of constructingmachinery is underpinnedby logic, mathematics andcomputer science.

A significant discovery in any oneof these disciplines could impacton the development of AI.

The big questions of “mental processes” tackled byAI are bound to a number of disciplines –psychology, philosophy, linguistics and neuroscience.

14

AI and Psychology

The objectives of AI and psychology overlap. Both aim to understand themental processes that underpin human and animal behaviour.Psychologists in the late 1950s began to abandon the idea thatBehaviourism was the only scientific route to understanding humans.

Behaviourists believe that explanations for human andanimal behaviour should not appeal to unobserved“mental entities”, but rather concentrate on what wecan be sure of: observations of behaviour.

Instead of restricting the object ofstudy to stimulus–responserelationships, those who abandonedBehaviourism began to consider internal“mentalistic” processes, such asmemory, learning and reasoning, as avalid set of concepts for explaining whyhumans act intelligently.

15

Cognitive Psychology

Around the same time, the idea that the computer could act as a modelof thought was gaining popularity. Putting these two concepts togethernaturally suggests an approach to psychology based on a computationaltheory of mind.

By the end of the 1960s, cognitive psychology had emerged as a branchof psychology concerned with explaining cognitive function ininformation-processing terms, and ultimately relying on the computer asa metaphor for cognition.

… within 10 years,psychological theorieswill take the form ofcomputer programs.

In 1957, Herbert Simon(1916–2001), an AI pioneer,made the prediction …

16

Cognitive Science

It is clear that AI and cognitive psychology have a great deal of commoninterest.

AI sits alongside cognitive psychology at thecore of an interdisciplinary approach tounderstanding intelligent activity.

The concepts in this book thereforerightfully fall within the remit ofcognitive science, as well as AI.

This has naturally led to a commonpursuit known as cognitive science.

17

AI and Philosophy

Some of the fundamental questions asked by AI have been the hardstuff of philosophers for thousands of years. AI is perhaps unique in thesciences. It has an intimate and reciprocal relationship with philosophy.

In one survey, AI researcherswere asked which disciplinethey felt most closely tied to.

The most frequent answer was philosophy.

18

The Mind-Body Problem

The mind-body problem dates back to René Descartes (1596–1650),who argued that there must be a fundamental difference between themental realm and the physical realm. For Descartes, man was alone inhis possession of a mental faculty – animals were mere beasts lackingany mental life.

But in the case of man, how canthe physical body be affected byprocesses occurring in the non-physical mental realm?

This is an age-oldconundrum …

19

AI informs modern discussions of the mind-body problem by proposingthe computer metaphor, which draws a parallel between the relationshipof programs to computers and minds to brains.

In a similar way, our mindcan affect our body.

Computer programs, like minds, haveno physical mass yet patently have acausal connection to the physicalcomputer executing the program.

Computer programs require acomputer to manifest themselves– just as a mind requires a brain.

20

Ontology and Hermeneutics

Attempts to equip machines with knowledge require one to makeontological assumptions. Ontology is the branch of philosophyconcerned with the kinds of things that exist. AI projects, lasting tens ofyears, have attempted to distil common-sense knowledge intocomputers.

But recently, these criticisms have shaped new approaches to lookingat cognition, and have had a positive influence on AI. We will return tothis later.

Insights originating from the branch ofcontinental philosophy known asHermeneutics have vehemently criticizedthe very possibility of formalizing mentalprocesses in this way …

To do so, the designers have to decide onthe “kind of things” a machine must knowin order to make sense of the world.

21

A Positive Start

The term Artificial Intelligence was coined at a small conference atDartmouth College, New Hampshire, in 1956. Some of its key figuresgathered to discuss the following hypothesis ...

This hypothesis has been subject to intense research ever since.Many of those attending the conference went on to be pivotal in thestudy of AI.

22

Optimism and Bold Claims

The Dartmouth conference ran for two months. Two attendants inparticular, Allen Newell and Herbert Simon, provoked much discussionby claiming ...

AI has always provoked great interest. The possibility of thinkingmachines has been a mainstay of science fiction. This is partly a resultof our fascination with the limits of technology and partly due toenthusiastic AI researchers.

We have invented a computer programcapable of thinking non-numerically …

And thereby solved the venerablemind-body problem.

This was perhaps the first of a longlist of bold and enthusiastic claimsthat litter the history of AI.

23

One common criticism of AI is its unashamed self-publicity, as T. Roszakcomplained in the New Scientist in 1986: “AI’s record of barefaced public

deception is unparalleled in annals of academic study.”

This statement is still dubious nearly 50 years later. Can machines reallythink? As we will see later, this is an important question, but it is riddledwith conceptual problems. However, a strong case can be made for theexistence of machines that can learn and create.

It is not my aim to surprise you or shock you– but … there are now in the world machinesthat can think, that learn and create.

In 1957, Herbert Simon arguedthat machines could think …

24

Intelligence and Cognition

So what exactly is intelligence, and how do we decide when somethingis artificial, rather than the real thing? Neither of these questions admitsprecise definition, which makes Artificial Intelligence an unfortunatename for a branch of science. On the concept of intelligence, A.S. Rebernoted in 1995: “Few concepts in psychology have received more

devoted attention and few have resisted classification so thoroughly.”

In the context of AI, intelligent is best takento mean “exhibiting interesting behaviour”.

Interesting behaviourcan be found in ants,termites, fish and mostother animals …

But these animals are not considered intelligentin the everyday sense of the word.

25

The relationship between behaviour and intelligence is rife withproblems. To illustrate these problems, we will consider perhaps the firstmilestone in autonomous robotics.

Intelligence is the computational part of theability to achieve goals in the world. Varyingkinds and degrees of intelligence occur inpeople, many animals and some machines.

Humans undoubtedly exhibit manyinteresting behaviours not observed inother organisms – for example, language.

So there are varying degrees ofintelligence, with humans sittingat the “high intelligence” end ofthe spectrum.

26

Mimicry of Life

During the 1950s in Bristol, south-west England, W. Grey Walterpioneered the construction of autonomous robots. Walter carried out hisinfluential work long before the availability of digital computers. He wasinterested in Cybernetics – the study of the range of possible behavioursof animals and machines.

Walter was interested in the “mimicry of life” and built robots thatcontinue to draw interest today. Using very basic materials, such ascogs from gas meters, Walter constructed a series of mobile robots thatresembled tortoises.

This means that the same principles canapply to all three, even though they mightbe made from very different materials.

Cybernetics rests on the assumptionthat the laws that govern thecontrol of humans, animals andmachines are universal.

27

These robots were autonomous. There was no human intervention orcontrol governing their behaviour. Walter’s robots had three wheels andwere surrounded by a shell that acted as a bump detector.

Using two motors to control the lead wheel, one for steering, and one forpropulsion, the robot would seek light. However, when faced withextreme brightness, part of the robot’s design made it avoid the sourceof the light.

As well as detecting collisions with objects,the tortoise also had a light sensor …

I am designed to beattracted to light.

28

Complex Behaviour

Walter reported that one of his creatures, Elsie, exhibited unpredictablebehaviour. For example, as part of Elsie’s environment, Walterintroduced a hutch containing a bright light and a re-charging station.

She would now enter what appeared to be a dimly lit hutch and re-charge herself. When full power was restored to the battery, fullsensitivity would return, and Elsie would dash out of the hutch and carryon as before.

After darting around in an animal-likefashion, Elsie’s on-board battery would rundown, and her usual behaviour of avoidingthe brightly lit hutch would change.

With fading batterypower, my sensitivity tolight would diminish.

29

Is Elsie Intelligent?

Walter’s creatures were very simple by modern standards, yet they shedlight on issues confronting contemporary robotics by illustrating howcomplex behaviour can arise from simple machines. There was no wayWalter could predict the exact behaviour of his robots.

But the capabilities of Elsie are a far cry from what we consider “real”intelligence. Importantly, Elsie has a lot in common with the famoushorse known as Clever Hans.

Elsie’s behaviour depends too muchon the environment and factors suchas fading battery power.

I could certainly achieve goalsin the world, since I couldsustain my own battery power.

30

Clever Hans: A Cautionary Tale

Clever Hans was a horse famously taught to do arithmetic by his trainer,Wilhelm von Osten. Hans would tap out the correct answer to a problemwith his hoof, to the amazement of the onlooking crowd, and onlyoccasionally make a mistake. Scientific experts supported his trainer’sclaims: Hans really could do arithmetic. But one expert noticed thatHans was making mistakes when von Osten himself didn’t know theanswer. Hans’s cover was blown.

Making a “Clever Hans error” means mistakenly attributing a capacity toan agent when in fact the capacity is supplied by the environment – inthis case, an arithmetically competent human.

The horse is being givencues by von Osten thatindicate when it shouldstop tapping its hoof.

31

Believers of Clever Hans mistakenly ascribed von Osten’s intelligence tothe horse. Similar criticisms have been levelled at W. Grey Walter’srobotic tortoises.

This illustrates the problem of ascribing a capacity to an agent solely onthe basis of its behaviour.How can AI construct intelligent machines, when intelligent action is sointimately related to the environment? The majority of AI researchhas side-stepped this problem in two ways. First, by focusing oncognition in agents detached from the complexities introduced by real-world environments. Second, AI mainly concerns itself with studyinginternal cognitive processes, rather than external behaviour.

Hans cannot count, and Elsie hasno desire to maintain her power.

Both Hans and Elsie appear to behave intelligently,but neither of them actually possesses the capacitythat their behaviour suggests.

The environment they inhabit hasbeen carefully designed by Walterto elicit the desired behaviour.