Mehta Ankit Chandrakant asked:
AI is the science of making machines do things that would require intelligence if done by men. It includes reasoning, learning, planning, speech recognition, vision, and language understanding. These machines are being used today in a wide variety of applications, such as monitoring credit card fraud, making autonomous decisions on space missions, watching for attacks from computer network hackers, diagnosing faults in aircraft, enabling human–machine speech interfaces, and making the characters in a video game behave in a more human-like way.
The main unifying theme is the idea of an intelligent agent. We define the AI as the study of agents that receive percepts from the environment and perform actions. Each such agent implements a function that maps percepts sequences to actions, and we cover different ways to represent these functions, such as real-time conditional planners, neural networks, and decision theoretic systems. We treat robotics and vision as occurring in the service of achieving goals.
Eliza, the program was able to converse about any subject, because it stored subject information in data banks.
Keywords: Turing test – Intelligent agents - Neural Networks – Genetic Programming – Planning – Fuzzy Logic – Robotics – Pattern recognition - Natural language processing – Deep blue - Eliza - video clip.
Index:
Introduction
Turing test
Classification
Intelligent agents
What belongs to AI?
Applications
Chess and AI
Computer Vs. Human Brain
Fuzzy-Logic and AI
Eliza
Conclusion (AI present & future)
References
1. Introduction
After WWII, a number of people independently started to work on intelligent machines. The English mathematician Alan Turing gave a lecture on it in 1947. I would compare attempts to create AI with man’s historical attempts at flight.
2. Turing test (Article of Computational Intelligence: 1950)
He argued that if the machine could successfully pretend to be human to a knowledgeable observer then you certainly should consider it intelligent. The observer could interact with the machine and a human by teletype (to avoid requiring that the machine imitate the appearance or voice of the person), and the human would try to persuade the observer that it was human and the machine would try to fool the observer.
3. Classification
This is a discipline with two strands: science and engineering. The scientific strand attempts to understand the requirements for, and mechanisms enabling, intelligence of various kinds in humans, other animals and information processing machines and robots. The engineering strand attempts to apply such knowledge in designing useful new kinds of machines and helping us to deal more effectively with natural intelligence, e.g. in education and therapy.
Bottom-up theorists believe the best way to achieve artificial intelligence is to build electronic replicas of the human brain’s complex network of neurons, while the top-down approach attempts to mimic the brain’s behavior with computer programs. Moreover there is a lot of difference between the AI simulated system and a program with the large database. This is discussed later on under the topic of chess-AI.
4. Intelligent Agents
An agent is anything that can be viewed as perceiving its environment through sensors and acting upon that environment through actuators. An agent’s percept sequence is the complete history of everything the agent has ever perceived.
A rational agent is one that does the right thing. A Performance measures embodies the criterion of a success of an agent’s behavior. When an agent is plunked down in an environment, it generates a sequence of actions according to the percepts it receives.
The Nature Of Environments
Specifying environment (PEAS – Performance, Environment, Actuators, Sensors) under the heading of the task environment we group the performance measure, the environment, and the agent’s actuators and sensors.
E.g., an automated taxi driver.
Agent Type Performance
Measure Environment Actuators Sensors
Taxi driver Safe, fast, legal, comfortable trip, maximize profits Roads, other traffic, pedestrians, customers Steering, accelerator, brake, signal, horn, display Cameras, sonar, speedometer, GPS, odometer, accelerometer, engine sensors, keyboard
It is better to design performance measures according to what one wants in the environment, rather than according to how one think the agent should behave
More specifically, the learning ability needs to be autonomous, goal-directed, and highly adaptive:
Autonomous — Learning occurs both automatically, through exposure to sense data (unsupervised), and through bi-directional interaction with the environment, including exploration/ experimentation (self-supervised).
Goal-directed – Learning is directed (autonomously) towards achieving varying and novel goals and sub-goals — be they ‘hard-wired’, externally specified, or self-generated. Goal-directedness also implies very selective learning and data acquisition (from a massively data-rich, noisy, complex environment).
Adaptive – Learning is cumulative, integrative, and contextual and adjusts to changing goals and environments. General adaptivity not only copes with gradual changes, but also seeds and facilitates the acquisition of totally novel abilities.
5. What Belongs to Artificial Intelligence
Neural Networks
Artificial Neural Networks, often just called Neural Networks (NN), are modelled on the human brain. The internal structure of the network, composed of a small number of artificial neurons, implies that the information learnt is not perfect. Artificial Neural Networks have been used successfully in visual pattern recognition, even human faces and complex industrial components can be differentiated. Artificial Neural Networks have been used in speech recognition system to decipher audible language.The technique used is that of a highly parallel network of simple processing elements. Each element has some similarity with animal nerve or brain cells called neurons
Genetic Programming
Genetic programming is an excellent way of evolving algorithms that will map data to a given result when no set formula is known. Mathematicians/programmers could normally find algorithms to deal with a problem with 5 or so variables, but when the problem increases to 10, 20, 50 variables the problem becomes close to impossible to solve. Briefly, how a GP-powered program works is that a series of randomly generated expression trees are generated that represent various formulas. These trees are then tested against the data, poor ones discarded, good ones kept and breed. Mutation, crossover, and all of the elements in genetic algorithms are used to breed the ‘highest-fitness’ tree for the given problem. At best, this will perfectly match the variables to the answer, other times it will generate an answer very close to the wanted answer.
Planning, problem solving, automatic design:
Planning involves finding a sequence of actions that can lead from the current state, to the goal state. Given a complex problem and a collection of resources, constraints and evaluation criteria create a solution which meets the constraints and does well or is optimal according to the criteria, or if that cannot be done propose some good alternatives.
Machine Learning
Machine learning is becoming increasingly popular, and equally important. People realise that it is theoretically much easier to get a machine to learn something from facts, rather than have to spend time teaching it explicitly. The quality of the learning algorithm is of course a major factor!
Constraint Satisfaction
Here, the problem is modeled as a set of variables, which can be assigned particular values. Different types of constraints are set-up on these variables (equality, numerical constraints), in order to specify the requirements for the problem. A search is then performed on the variables, in order to find the potential solutions. There are many nifty tricks involved to partly resolve constraints in order to guide the search more efficiently (this is called a heuristic search). The problems solved can also be a combinatorial optimisation, where a particular solution has a better value than another, and the best needs to be found. The class of problems usually solved is NP-complete, where the complexity increases exponentially as the problem size increases linearly.
Search and Optimization
There are many kinds of searches, the simplest of which involve trying out all the solutions in a particular order. The entire set of possible solutions is called the search space.
Decision Tree Learning
A decision tree is a structure that allows learning of opinions (e.g. good or bad) about objects based on their attributes (length, colour…). Given a series of examples, the learning algorithm can build a decision tree that will be able of classifying new examples. If the new examples are handled correctly, nothing is done. Otherwise, the structure of the tree is modified until the correct results are displayed. The challenge is getting the algorithm to perform well on very large sets of data, handling errors in values (noise), and determining the optimal fit of the tree to the training and test data.
Data Mining
This is the process of extracting useful rules from very large sets of data. Data Mining is a term used to describe the process whereby software tools examine a company data base in order to locate information which may have complex parameter connectivity. Such information would normally be inaccessible to the human expert due to the enormous quantity of data and combinatorial tests which would have to be performed. A simple example may be a data base of company products and parameters which describe their applicability to various sectors of the market.
Bayesian Networks
Bayesian Networks models the relationship between variables. This is called conditional dependence: a state of a variable may depend on many others. This can be represented as a graph, and there’s a clever algorithm to estimate the probability of unknown events given existing knowledge. Admittedly, one common complaint against this approach relates to the design; it can be very tedious to model such networks. As such, learning the structure and the inference between variables seems like an appealing option.
Artificial Life
Artificial Life (A-Life) is the study of artificial or computer based systems, which exhibit life like behavior. Computer simulations of individual agents or populations of agents can be used to investigate many of the properties of living systems. In some cases, mechanically constructed agents are provided with basic functionality and allowed to interact with real environments. This is a very popular aspect of Artificial Intelligence, which involves modeling and mimicking living systems. This includes anthills, wasp nests, larger forests, towns and cities. To date, very complex and interesting systems have been created by a multitude of very simple entities. For example many ants programmed by very small programs would potentially create an entire system with signs of emergent intelligence.
6. Applications
Robotics
The main aspect of robotics today is mobility. This can be done by learning the task in a virtual simulation, and then applying it to the real robot. If specific conditions of training are respected, the problem has a high probability of working in real life, but this is no guarantee. In practice when moving robotic arms, the arm has a few movement possibilities: the shoulder allows rotations according to two axes, and the elbow also allows two basic rotations. Each of these possibilities is called one degree of freedom. Usually, one controller is assigned to provide movement for one DOF. The task at hand is to learn the optimal combination of controllers, where they can successfully cooperate to perform a given task.
Pattern Recognition
Pattern recognition involves determining the characteristics in specific samples and sorting them into classes; a process called classification. This is usually done with Machine Learning techniques, allowing the system to adapt to the data given to it. It can be applied to detecting single words in speech, recognizing voices, sorting scanned objects by type and filtering out unwanted pictures (among many others). In practice, a way of doing this is to represent the sample as a set of features (e.g. for a sound: pitch, volume, timbre, smoothness). A training set is then created: i.e. a set of samples where the result is known (e.g. for ****** recognition: Fred has green eyes and brown hair, Henry has blue eyes and blond hair). The learning mechanism can then learn to associate the features with the known types of sound or image. Depending on the representation, more or less samples are required. With symbolic representations, small numbers of examples are usually required, whereas for fuzzy learning (in neural networks for example) larger training sets are needed.
Natural language processing
It includes production and interpretation of spoken and written language, whether hand-written, printed, or electronic throughout (e.g. email). One of the first approaches was symbolic, assigning semantic meaning to each word (verb, noun, adjective). The basic structure of valid sentences would have to be defined manually, and a search would be performed to match the template with the current sentence. A lot of time needed to be spent resolving ambiguous sentences, and getting the person and tenses of the verbs to match. If the programmer spends enough time creating the sentence templates, the results would be fairly encouraging. But this monotonous task needs to be repeated for new sentence constructs and new languages all together.
A very recent approach is to use statistical analysis of the text. In essence, large parts of books are processed and learning algorithms attempt to extract the rules and patterns. This requires a smarter approach, taking more time to design, but it results in a more flexible program.
Frames
On method that many programs use to represent knowledge are frames. Pioneered by Marvin Minsky, frame theory revolves around packets of information. For example, say the situation was a birthday party. A computer could call on its birthday frame, and use the information contained in the frame, to apply to the situation. The computer knows that there is usually cake and presents because of the information contained in the knowledge frame. Frames can also overlap, or contain sub-frames. The use of frames also allows the computer to add knowledge. Although not embraced by all AI developers, frames have been used in comprehension programs such as Sam.
AI in medicine, including interpretation of medical images, diagnosis, expert systems to aid GPs, monitoring and control in intensive care units, design of prosthetics, design of drugs, intelligent tutoring systems for various aspects of medicine.
AI in many aspects of engineering: fault diagnosis, intelligent control systems, intelligent manufacturing systems, intelligent design aids, integrated systems for sales, design, production, maintenance, expert configuration tools (e.g. ensuring sales staff don’t sell system that won’t work). AI in software engineering includes work on program synthesis, verification, debugging, testing and monitoring of software.
AI in education: including various kinds of intelligent tutoring systems and student management systems. Particular applications might include diagnosis of a student’s knowledge gaps, various kinds of drill and practice tutors, automatic marking of programming assignments and essays, etc.
AI in entertainment: increasingly AI is being used in computer games and in systems for generating and controlling synthetic characters either for textual interaction or generating films with cartoon characters or interactive avatars in virtual worlds.
AI in biology: there are many hard problems in biology where more or less intelligent computer-based systems are being developed, e.g. analysis of DNA, prediction of folded structure of complex molecules, prediction, modeling many biological processes, evolution, development of embryos, behaviors of various organisms.
Architecture, urban design, traffic management: tools to help solve design problems involving multiple constraints, helping to predict the behaviors of people in new environments, tools to analyze patterns in observed phenomena.
Literature, art and music: identification of authors, modeling of processes of generation and appreciation, teaching applications.
Crime prevention and detection: e.g. detection of forgeries, learning to detect evidence of crooked police officers, software to monitor internet transactions, helping to plan police operations, searching police databases for evidence that crimes are committed by the same person, etc.
Space: control of space vehicles and autonomous robots too far from earth to be directly manipulated by humans on earth, because of transmission delays.
Military activities: Various paradigms in AI have been successfully applied in the military field. For example, using an EA (evolutionary algorithm) to evolve algorithms to detect targets given radar/FLIR data, or neural networks differentiating between mines and rocks given sonar data in a submarine.
7. Chess and AI
Deep Blue does not use AI. Then how is AI – deep blue related?
AI-based game playing programs combine intelligence with entertainment. World-champion chess playing programs can see ahead twenty plus moves in advance for each move they make. In addition, the programs have an ability to get progressably better over time because of the ability to learn. Chess programs do not play chess as humans do. In three minutes, Deep Thought (a master program) considers 126 million moves, while human chessmaster on average considers less than 2 moves. The next move comes from exhaustive searches into all moves, and the consequences of the moves based on prior learning. Chess programs, running on Cray super computers have attained a rating of 2600 (senior master), in the range of Gary Kasparov, the Russian world champion.
DEEP BLUE: First of all, this year Deep Blue will be running on a faster system - the latest version of the SP -, which uses 30 P2SC or Power Two Super Chip processors. Last year, Deep Blue averaged about 100 million chess positions per second. This year Deep Blue will work about twice as quickly - that is, 200 million chess positions per second. Incidentally, Garry Kasparov can evaluate approximately three positions per second.
8. Fuzzy Logic and AI
It is often said that computers are too logical and that they can only deal in true or false, yes or no etc. However, Fuzzy Logic allows a computer to deal in everyday human language and actually process terms such as probably, unlikely, quite near etc. Such terms can take their place in computations, allowing the computer to arrive at verifiable results from fuzzy inputs. Another type of fuzzy information is stored in the famous Neural Networks. This is known as a neuro-fuzziness. The information inside an artificial neural network is usually imprecise, due to the weighted connection between neurons (called synapses).
Fuzzy representations have increased in popularity, due to the increased capabilities of computers: more processing power is usually required to create such rules, and interpreting them generally also requires a bit more time. The preferred languages for this type of representation are usually procedural like C, C++ or Pascal.
9. Brain Vs. Computer
A collection of simple cells can lead to thought, action, and consciousness or, in other words, that brains cause minds. Even thought a computer is a million times faster in raw switching speed, the brain ends up being 100,000 times faster at what it does.
Computer Human brain
Computational units 1 CPU, 108 gates 1011 neurons
Storage units 1010 BITS RAM 1011 neurons
1011 bits disk 1014 synapses
Cycle time 10-9 sec 10-3 sec
Bandwidth 1010 bits/sec 1014 bits/sec
Memory updates/sec 109 1014
10. ELIZA
Eliza, Joseph Wiezbaum’s result of trying to make a program converse in English amazed people when it appeared in mid 1960’s. The program was able to converse about any subject, because it stored subject information in data banks. Another feature of Eliza was its ability it picked up speech patterns
Conclusions:
Finally, we come up with many conclusions with regards to future and present of Artificial Intelligence. AI is fascinating, and intelligent computers are clearly more useful than unintelligent computers, so why worry? AI has made possible new applications such as speech recognition systems, inventory control systems, surveillance systems, robots and search engines.
Finally, it seems likely that a large-scale success in AI—the creation of human-level intelligence and beyond—would change the lives of majority of the humankind. The very nature of our work and play would be altered, as would our view of intelligence, consciousness, and the future of destiny of the human race. At this level, AI systems could pose a more direct threat to human autonomy, freedom, and even survival. After all silicon is cheaper than human life.
In conclusion, we see that AI has made great progress in its short history, but the final sentence of Alan Turing’s essay on Computing Machinery and Intelligence is still valid today:
“We can see only a short distance ahead, but we can see that much remains to be done.”
.
