|
features post Articles
(Articles)
online chat server:
irc.xor.cx
channel:
#neworderrandom article
quotable quotes What 'football' are you talking about? proper Euro stuff of wimpish American 'lets wear padding in case we get hurt and cry' stuff? munkt0n
|
A. I. - Can Computers Think?
@ Out of the box
Sep 23 2006, 08:52 (UTC+0) | C0B01 writes: To begin with I feel it is imperative to clear up a miss conception of Artificial Intelligence. AI is in fact not about the creation of artificially intelligent machines, it is the field of creating machines that are able to think rationally. Given sensory inputs the artefact is capable of producing a relative, rational response. A specific chain of events that ended in the result of the output; an event that may not have been programmed specifically for, in contrast to the single lined approach of traditional do-this do-that computational methods. So straight off, can computers think? My answer, in contrast to all those I asked is, “yes they can”.
|===================================================|
| AUTHOR : C0B01 |
| E-Mail : C0B01(at)hotmail(dot)com |
| Title : Can Computers Think? |
| Version : 1.1 |
| Word Count : 4959 |
| Original Release : December 2005 |
| Last Updated : May 2006 |
|===================================================|
Can Computers Think?
CONENTS
1 Introduction
1.1 What is Artificial Intelligence?
1.1.1 A View on AI
1.1.2 How do we Think?
1.1.32 The Answer to the Question
2 Languages & Symbols
2.1 What is a Language?
2.1.1 How do Languages work?
3.1.2 DNA as a Language
3 The Ethics of AI
3.1 Vernor Vinge
3.2 Current Capabilities
3.2.1 Computers at present
3.2.2 Humans at present
3.2.3 The new sccience
3.2.4 The Technnological Singularity
4 Machine Learning
4.1 Ways to Learn
4.2 Machines that Play Chess
4.2.1 Game Trees
4.2.2 Neural Networks
5 A Brief Look into the Future
6 Natural Language Processing
7 In Conclusion
1 Introduction
1.1 What is Artificial Intelligence?
1.1.1 What really is AI?
To begin with I feel it is imperative to clear up a miss conception of Artificial Intelligence. AI is in fact not about the creation of artificially intelligent machines, it is the field of creating machines that are able to think rationally. Given sensory inputs the artefact is capable of producing a relative, rational response. A specific chain of events that ended in the result of the output; an event that may not have been programmed specifically for, in contrast to the single lined approach of traditional do-this do-that computational methods. So straight off, can computers think? My answer, in contrast to all those I asked is, “yes they can”.
1.1.2 A View on A.I.
“To create an artificial being has been a dream of man since the birth of science, not merely the beginning of the modern age, when our forbearers astonished the world with our first thinking machines, primitive monsters that could play chess, how far we have come. The artificial being is a reality of perfect simulacra, articulated in limb, articulated in speech and not lacking in human response. ”
- A.I. ‘Stephen Spielberg’
Our insatiable fascination with the ideas of creating artificial life has resulted in great efforts and many resources being diverted into further research into the field of AI. Many see AI as one of the newest of mans sciences, but its roots go far beyond the coining of its name back in 1956.
1.1.3 How do we really Think?
Before we can even begin to understand how we can make a machine appear intelligent, we need to first try and understand how humans and animals think. Analysing the ways in which we solve problems and actively thinking ‘how we think’; documenting each stage of reasoning that went behind our decision. Taking a problem and going through it stage by stage in a logical manner applying governing rules to produce a foreseen outcome. The logic behind this being, you need to know what your goal is in order for you to carry out actions which will inevitably end in the desired result. Your hungry… in your head you can visualise a possible situation in the future in which you eating - this is your goal. In order for you to get to this given goal, you as an agent must perform a series of actions which resulting reactions will result in the accomplishment of the goal. You can not do anything intelligent if it was unintentional. An example would be; you can’t stop feeling hungry if you do not realise you want to be eating something. If you never had the thought to be eating something, you never would eat; hence death would be a certain outcome – Which Darwin’s theory of natural selection proves it is the reason why all autonomous agents in our world today, human to insect share the same characteristics; of controlled mobility, proactivity (goal orientated), reproduction, reactivity, autonomy, temporally continuous, and are adaptive to their environment.
Note: An environment is organism specific. That is to say an organism that has over time adapted well to its environment will no necessarily adapt well to a different environment. Only those capable of adapting will live to pass on there genetic code.
With these defined characteristics, we are now for the first time able to centralise our efforts towards the completion of these goals. In theory, if we are able to recreate these characteristics in an inanimate object, the question lye’s “would we therefore be able to define that object as now being a thinking, intelligent machine?”
1.1.4 The Answer to the Question
While doing my research for this paper I found it a helpful insight asking others who where unfamiliar to the field the question “can computers think?” to see their reaction. Not surprisingly all said “no” straight away and where confused by the questions logic. “Surly a computer can’t think? That’s what living things do; a computer isn’t a living thing, therefore it can’t possibly think”. In each case I tried to explain that there is a difference between the act of ‘thinking’ and consciousness. Rational thinking, that is, given inputs and applying known logic to them produce the rational output. That is – I know this is true, so if this happens and this implies that, then that is also true. Consciousness on the other hand as we as humans all know can make us think irrationally. A computer ‘thinking’ rationally will always produce the right output so long as the relations are known. It is not possible for a computer to perform irrationally unless it has bad data or incorrect logic. But maybe this is why some humans act irrationally…
2 Languages and Symbols
2.1 What is a language?
2.1.1 How do languages work?
Moving onto the side of thinking that we should all be familiar with - natural language. I’ve always perceived this as an amazing question, “how would we think f we didn’t have a language?” We think to ourselves constantly in our own language; that voice in our head’. Be that in English, Welsh or whatever country you where brought up in. All languages have one thing in common with each other - they are all representations of real life objects or things. Combinations of symbols (as in text) or as sound waves (changes in atmospheric pressure) which link together in specific ways to form meanings.
Looking at these links; a word is a string of characters, but not all strings of characters are words. The same as in a sentence is a concatenated string of words, but not all concatenations of words are sentences. These complex patterns of symbols once processed to their real life representations, in a ‘virtual world’ inside our brain enable us to visualise the problem. Again applying known knowledge we can foresee multiple outcomes and then choose which path to take. I am unfortunately unable to recall the originator for the following information but I have read that some believe that consciousness and free will is just our perception of the choice phase. And that free will isn’t really that free. Your subconscious does a lot of thinking without your knowledge, and only the outputs of this thinking, the possible choices, are perceived by our consciousness, which then chooses which action to take- known as the choice phase.
2.1.2 DNA as a Language
Keeping with the idea of symbols and relationships, life’s instructions DNA, consists of four polly-nucleon types, which contains all the genetic information needed to be able to reproduce the original organism to which it belonged. Looking at this with a more computational view, DNA is the four symbol code that constructs life. In modern computers we use a two symbol language that is used to encode our programs. Future tri state processors will be able to move us forward to a three symbol language in computing, getting us much closer to the apparent capabilities of the four symbol language of DNA. Maybe some day we will be able to program DNA computers, it is estimated that a single sunflower seed has the data storage capacity of 100,000 text books. If we as humanity are capable of fully unlocking the keys to DNA, and are able to manipulate it to our will who knows what the future may hold…
3 The Ethics of AI
3.1 Vernor Vinge
Vernor Vinge is known to have said “Within thirty years, we will have the technological means to create superhuman intelligence. Shortly after, the human era will be ended.” This is a somewhat extreme view, but is still considered by some to be valid. It is also known that technological progress in the field is growing exponentially at present. Every other technology has demonstrated an S-shaped curve, where the exponential growth eventually tapers off. But so far there has been no indication of a near future slow down, but obviously it is unrealistic to believe that this exponential increase will continue indefinitely. Humans on the other hand with their complex genetic evolution do not measure well against the advancements in hardware and software on the long term scale in relation to physical evolution.
3.2 Current Capabilities
3.2.1 Computers at present
One of the most advanced super-computer---driven agents in our world today works at around 10 million calculations per cycle. This may seem a lot, but in comparison to the human brain estimated at around 100 trillion calculations per cycle, the computers have a lot of catching up to do. But with new visions of revolutionary hardware, such as optic processors – calculating at the speed of light - on the horizon, who knows how long it will be before the machines match and even exceed our capacity.
3.2.2 Humans at present
The human brain is a massively parallel system, with no single part taking precedence over the rest. Instead brain activity is distributed across different areas of the brains hemispheres, with most activity located to the outer regions. These simultaneous reactions produce the final thought. Whereas in traditional computing the central processing unit applies a serial approach, with each process being handled in the same area - one instruction after another. This bottle neck in effect slows down the computers ability to calculate and execute programs. With newer technology, parallel and distributed systems we are coming closer to having the technology that will most certainly be required for future ‘thinking’ agents.
An organic processor such as the human brain is capable of sustaining limited damage with minimal loss to overall operation. An example would be of the case of a 30 year old American who tried to commit suicide while intoxicated on narcotics. While under the drugs influence the man embedded 20 six inch nails into his skull with a gas powered nail gun. The man survived – with little visible damage to the subjects mental abilities. This kind of durability is in contrast to that of the delicate construct of a computer hardware. A single fault in a component can bring a whole system down. Anything from a overheated capacitor that’s exploded over the load, or a simple short across the tracking can completely halt a systems ability to operate.
3.2.3 The new Science
The new science of AI has also brought up a moral debate over the idea of thinking machines. At what level of thinking does one arrive at conscious thought? Is it really possible for a machine to become self-aware? And if so, what implications would that have on us as a species? However science fiction these questions may seem we must face up to them as if they were a reality. And treat the subject with due respect. It was not 300 years ago that the idea of being able to talk to someone and see them and be on the other side of the world at the same time was seen as such a fantasy.
3.2.4 The Technological Singularity
These serious questions have resulted in theories such as the technological singularity or ‘intelligence explosion’ to be created. The theory focuses on the idea that if man was completely successful in creating a truly intelligent machine, then that machine would be intelligent enough to design an even more intelligent one than itself, and so the explosion begins. A near infinite exponential rise in intelligence would arrive, with its end far out of the reach of any human eyes.
4 Machine Learning
4.1 Ways to learn
A major part of thinking and the ability to think involves the act of learning. We all should know that the best way to learn is to hypothesize beforehand, before you have any distinguishable knowledge to prove a theory either correct or otherwise. This enables you as an agent to see, even if you don’t fully understand, the relations that occur in a virtual world (the one inside your head as you are thinking). And then use this knowledge in the aid of future problem solving. There is a search technique known as the ‘current-best-hypothesis search’ originally compiled in 1843, but its believed that the algorithm fundamentals have been around for much longer. The current-best-hypothesis search uses a system of classification to distinguish whether the subject matter would fall into the hypothesis’ explanation or fall in the outer ‘unexplained’ region – an area which the hypothesis has no answer, other than it is not the answer to the hypothesis. The below diagram visualises this categorisation nicely.
With the single ‘best’ hypothesis being the centralised greyed out area consisting of positive results that back up the hypothesis’ claims (a). Later when more processing has been involved and new information has been received about the subject matter an inconsistency (b), shown as the positive anomaly, proves that the current hypothesis is incorrect and does not fill all needs. So the solution is found in (c) where the hypothesis is expanded and generalized to include this new knowledge. On the other hand if an anomaly should occur actually within the hypothesis region itself, then the hypothesis must be specialised to take this into account, demonstrated in (d) and (e). With this constant updating of the hypothesis with the arrival of new data, a ‘current-best-hypothesis’ is always available, so at least some decision can be made with the limited knowledge.
The problem that this ‘learning technique’ is that it does not guarantee the derivation of the simplest hypothesis, just a working one. Which could be more costly in both memory requirements and time efficiency, rather than a simpler more effective route. With the current-best-hypothesis search technique a problem situation arises due to generalisation and specialisation of the hypothesis to a level where the altering of the hypothesis has not necessarily created the simplest of hypotheses and has created what will later evolve to be a unrecoverable problem.
Are our visions of robotic humanoids commuting with us in everyday lives, that resembling those scenes portrayed in science fiction stories such as the ‘Animatrix’ and ‘I Robot’ like those of our distant relations, whose minds viewed ideas such as flight as a replication of nature. Before science explained to us the laws of aerodynamics our efforts had all been aimed the wrong way. We can not fly like a bird, imitating nature, instead we have learned to glide through the air - on the wings of science. Early attempts at flight where met with no reward; but the idea was there – we wanted to fly through the air like a bird does. Of course we all know this isn’t possible, and the statement sounds almost rudimentary. But the point is that a plane flies, so too does a bird; but both do in very different ways. With this knowledge in mind it seems highly plausible to say that future strong AI systems will not in fact resemble those of our science fiction writers imaginations.
4.2 Machines that play Chess
Moving onwards to a more currently deployed use of AI, the game development side; in 1997 IBM’s chess program Deep Blue successfully out moved the world champion, Kasparov to take the victory. Deep Blue was a specially designed IBM server based ‘ration thinking’ program. Which had specially designed chips to aid in its enormous searches; Deep Blue had an opening book of around 4000 positions. (An ‘opening book’ is a database of ‘knowledge’ of possible start scenarios, these scenarios go many, many moves into the game – listing numerous possibilities of opponents moves ). Deep Blue also had Knowledge of 700,000 grandmaster games. By knowledge here I’m referring to the structured data the game tree provides. Deep Blue also had an evaluation function split into 8,000 parts, many of which where designed for specific situations occurring within the game. The Computer was also aided with the knowledge of all games ending with five or less piece positions, plus an incomplete selection of six piece end games.
Backing up the argument that computers are capable of thought, Kasparov said “he sometimes saw deep intelligence and creativity in the machines moves”. Going by an approach similar to the turning test, isn’t a main concept of AI the ability to have an interaction with a computer that is seemingly endless. In the game world, the world that is the chess board, surly deep blue has achieved a status to that of a fully interactive entity, at least limited to its own world. As we too (bar the few anomalies - astronauts!) spend our entire existence on Earth - our world, in a unimaginably vast universe. In respects to Deep Blue, any existence out of its world – the chess board could be seen upon as its universe. It is able to use its current situation, the pieces and their positions on the board, and with the implementation of previous knowledge, in the form of a game tree, uses different search techniques to come up with an appropriate move. It can learn by having nodes added for different moves and by altering the weights of the connections.
4.2.1 Game Trees
The game tree starts of at the root node, the node at which represents all the pieces on the board, set in their starting position. From this first node branches off every possible node that could take place. I.e. there will be say a node for the knight moving three up two left, another for the same knight moving three up and two to the right. And so on for each movable piece. You wouldn’t need a node at level one for a rook; as it is unable to move anywhere on the first move, same for bishops, the king and the queen. In fact only the knights and the pawns actually have legal moves on the first go. This factor of legal moves sets the size of the game tree – which can become very large indeed! Often an advanced chess program will search fifteen to twenty moves ahead into a game before it plays its move. Again going back to the initial question of thinking, the computer has its knowledge of the game, possible moves and outcomes, the same as what a human player might have stored in their brain. The human side of thinking evaluates and finds the best possible solution the same way as the computer does. Ignoring the black box functionality of the processes involved, the outcome is the same, a rational response aided by the desire to reaching a terminal state where you (you could be the computer) are the winner.
But how does it think (search)? The game tree can be searched in different ways. The program searches the tree using depth first search, starting on the left most node on the game tree, searches down each node in an attempt to find the best possible move, this is indicated by values anchored to the trees connections.
It is at this point I feel appropriate to mention the concept of minimax. In the game you must fully understand the concept of their being two, or possibly more players. These players need to be categorised as hostile (opponent) or allied. On your move you are not just trying to win, you are also trying to ensure your opponent (all hostiles) loose, for you to enable that win (for you and your allies). You also need to remember this goes the same for your opponent; he is trying to make you loose.
On your current move you must search for what moves you can do to that would give you a greater position in the game to enable a win, as well as taking into consideration the values affected by the positions of your opponent and his possible future moves. Taking both into consideration and going by the one which will maximise your chance (current score) and minimise your opponents chance (current score). This move would be seen as ‘the best possible move’ and should be the one executed.
4.2.2 Neural Networks
At present computers and humans are very close in abilities at playing at chess. Another game playing program TD-Gammon 1995, which in contrast to the somewhat traditional approach to the Deep Blue and similar chess program, TD-Gammon takes a different approach; using neural networks, a black box system in which the processes involved are unknown the system is able to learn. The program uses a system where it randomly generates the connection weights between the output units. These units are the resultant output from the hidden blackbox layer, which can be evaluated and scored accordingly. Process input is fed into the blackbox network with 198 input pins. For each position and for each colour on the board there are four inputs. With no starting knowledge TD-Gammon is capable of learning by playing against itself and learning from its experience to a level compared to that of a human seen as a strong intermediate player. Moreover than this, with the addition at the input layer of handcrafted features that add to the networks input representation, the neural network is capable of playing at levels that are exceptionally close to the abilities of the best human players.
It was also strangely observed that the network learned to play better by playing against itself rather than being trained against a human expert data set. The self teaching network, the one that is supposedly ‘thinking’ and ‘learning’ for itself was able to learn the game to near-complete knowledge of the games moves. This was in contrast to the one which was simply taught how to act in given situations by the human expert data set, which when compared didn’t perform as well. It is slightly worrying that the success of the TD-Gammon neural network is not fully understood, even by its creators.
Before we can look further into the question of can computers really think we need to first clarify what we define as being able to ‘think’. If we had a disembodied chess program, and we isolated the part of a disembodied human mind that governs the abilities to play chess, these two entities perform the same actions. We consider ourselves to be thinking when we play chess, so surly the correlation is obvious. And that it is arguable that the components and actions involved in the disembodied chess program would be seen as terminally problem solving, in other words thinking of a way to win. The medium in which this is taking place; be that in neurons and synapses or specially designed data storage and manipulation circuitry and components, surly does not effect the overall outcome - the ability to play chess.
5 A Brief look into the future
Theorising, if we are able to keep evolving AI techniques; search algorithms, data structures and abstraction of data types and are able to combine all elements then the final construction will provide vast benefits for its human users. – Imagine a intelligent security system, capable of effective voice recognition, which can distinguish different objects; has face recognition capabilities so as to differentiate person to person, and is able to call emergency services/other systems and deliver messages. With an alarm system like this there would be an end to false alarms where such things as motion-static / temperature variations etc… would set the primitive sensors off. The system would be able to know who was authorised and therefore allowed in the building and who was unknown and seen upon as hostile. If the face recognition was successful enough at characterising emotional state, i.e. it can tell whether the target is happy or in a state of fear, it would then be able to act accordingly. This system would be useful for immediate flagging of a hostile situation. Applications for such a system would be in installations such as chemical laboratories, nuclear reactors, munitions storage sites and other such high-security facilities.
This kind of implementation of AI brings us onto the idea of communicating with an agent via means other than that of the comparatively simple communications of a GUI. Although there are many areas, the ability to talk to a computer to give commands rather than the traditional keyboard input, brings us into the field of natural language processing.
6 Natural Language Processing
When we start looking into the area of natural language processing we are faced with a complex environment. Our languages are made up of phonemes, certain pitches of sounds grouped together. A solution may be to create a program that is capable of pattern matching the phonetics of a sentence and translating into a symbolic representation (text) which is much easier to manipulate and understand. This isn’t without its problems though… “wreck a nice beach you sing calm incense”. This nonsense sentence is phonetically the same as the correct vocalised original sentence “recognised speech using common sense”. Our brain is able to use many things in order for us to make sense of spoken word; such as the context in which it was spoken, who said it, and which word out of similar words would be the one used – “which witch is which witch?”. We of course all know that speech isn’t something you are born with. You are born with the capabilities of speech; but it is only after years of being bombarded with spoken word in varying situations, that you first start to understand the symbolic language. A baby learns by receiving audio input from the parents speech and by using the input from its vision at the same time is able to see the context in which those words are used, and eventually through different learning techniques, mainly reinforcement, be it positive or negative, is able to put the symbols into context. If I do this action, and mum shouts “no!” and I do this, and she keeps on shouting, which is something I don’t like (a non desired terminal state), so if I change my actions and do this, she stops shouting and gives me a sweetie. Here the child has learned what reaction is desired when another person says ‘no’, i.e. the relation in not doing something. You’ve learned the rule BeingNaughty equals MumShoutsNo implies !BeingNaughty. Natural language processing can be aided with the knowledge of typical sentence construction, words that more often than not group together. Using this predictive approach it is possible to search a database storing word combinations, which can be checked for the likelihood of the next word. It is practically helpful with deciding which word out of several phonetically similar words to use, and of course these words have totally different meanings, so their right use is important if the sentence is to be correct.
7 In Conclusion
The hardware industry is constantly finding newer, better ways of manipulating and storing data. It was only ten years ago that one single megabyte of RAM set you back over £100, with the modules consisting of multiple serial lines of silicon chips. Today, well over 500 times more RAM can be fitted onto a module less than a quarter of the size, for a fraction of the price! With these advancements in hardware, distributed and clustered systems, advancements in object orientated programming languages, and specialized AI languages such as Prolog and Scheme it is understandable to say that the future of AI is looking bright. Not only due to the fact that we are developing the technologies required but we are also due to other areas in technology are coming to the actual need of such systems. With data storage mediums becoming cheaper and cheaper and the use of recorded points of sale transactions becoming more frequent, huge amounts of raw data is being stored in specialised data warehouses. It would take an enormous task force of specially trained human data analysis experts to try and cover these amounts of data. Instead through the deployment of specialised AI agents mining data for useful information is now possible. This is of course is still not without its drawbacks, unrelated relations, or indirectly related relations may be discovered, and as to date are still supervised by a human user. A good example of a bad correlation is that of the rise of suicides in correlation to the rise in ice cream sales. It is known as a fact that as ice cream sales rise, so to does the suicide rate. This correlation to us obviously doesn’t make much sense on its own, but when you take other factors into consideration such as ice cream sales rise in the summer – when it’s hot. Which is a time of the year known for making most people feel generally happier. This psychologically have a negative effect on those already suffering from depression. When people around them are having fun, the people prone to suicide become more likely to become even more depressed and hence decide to end their life.
PostScript
==========
This paper is still work in progress, any ideas, contributions would be greatfully recieved.
-C0B01
-c0b01(at)hotmail(dot)com |
| Top of page
|
