Many of us probably do not realize just how involved computers are in our lives. But the reality is that computers can assist us in making travel arrangements, medical diagnosis and driving, with little of no human intervention. But will their intelligence ever surpass that of their makers?

Many people are quite confident that we will reach that goal. Hugh Loebner in coordination with Cambridge Center for Behavioral Studies has devised an annual award The Loebner Prize to systems who are willing to take the challenge of the "Turing Test". Each year the most successful system is awarded a prize of $2000.00 and a bronze metal. When the day comes the grand prize of $100,000.00 and a gold metal will be awarded to the first computer who successfully passes the "Turing Test". But exactly what will it take to pass the "Turing Test"?

"The Turing Test"

In 1950 Alan Turing defined intelligent behavior as the ability to perform all human task well enough to fool a human being. His test proposed that an interrogator would pose a series of intelligent and intellectual questions and tasks which would be performed by both a human and computer. If the computer was successful the interrogator would not be able to distinguish between the human answer and the computer's answer.

What would the science of Artificial Intelligence need to pass the Turing Test? If we review everything a computer would need to pass for humanlike we may come up with a list such as this (Russell and Norvig,1995,p. 5).

• Natural language processing
• Knowledge representation - The ability to store information.
• Automated reasoning - The ability to use stored information to answer questions and draw new conclusions.
• Machine Learning and Reasoning - The ability to adapt to new environments.
• Computer Vision - The ability to view objects and surroundings.
• Robotics - The ability to manipulate objects.

Natural Language Processing

Although speech is simple to us it involves many steps. A speaker must go through many steps to communicate an idea.

• The development of an idea to communicate.
• Choosing how the idea is to be communicated.
• Verbalizing the idea.

• The initial response to what the speaker has stated.
• Analysis of what the speaker has stated.
• Deciphering the meaning of what the speaker has stated.
• Deciding if the statement is true or false.

Several techniques have been devised to assist computer systems with natural language.

Sentence Parsing

The ability of a system to take apart a sentence and label the various parts as objects, nouns, adjectives, pronouns etc. This procedure assist the systems in defining and understanding the thoughts behind the communication.

The Use of Surrounding Words

The ability of a system to use the surrounding words to define words which have several meanings.

Models of Dialog

Models of dialog assist a system to define sentences which are otherwise incomplete. A obvious sentence to us such as "Dinner is ready" can be quite confusing to a computer system. The models of dialog technique would define the sentence as "The dinner is ready to be eaten". Without models of dialog the system would be at a loss. A computer system might think, the dinner is ready to do what?

Informal Rules of Conversation

Informal rules of conversation define to the computer system when a well defined answer is required and when a simple yes or no will suffice.

Background Knowledge of the Environment

We often do not realize how our historical knowledge of our surroundings assist us in every day life. Every understanding of human emotion and reactions has been learnt by events which we have lived through. Computer systems without this knowledge would not be able to interrupt such things as human emotions or body language.

Even though we have made great strides in determining the needs for computer systems to understand natural language, programming the systems to do so has proven to be quite a difficult task. Even when the goals mentioned above have been conquered there is still the issue of how we would program a system to understand sarcasm, slang or jokes. Natural language systems of today can only communicate on a very limited basis. An example of the limitations would be an instance when a computer system defined the phrase "Out of sight out of mind" as "Blind, Insane" (Gottinger and Weimann,1990, P. 64).

Knowledge Representation

The ability to store information. Currently the human brain has a much higher storage capacity than a computer system. The human brain can store 50 trillion bits of information while the a computer system can only store 1 trillion bits of information (Decker and Hirshfield, 1998, p. 303). Predictions have been made that if computers continue to progress at their current rate they will equal the storage capacity of the human brain by the year 2020 (Kurzwell, 1999, p. 203).

The storage capacity of a system greatly depends on the knowledge representation language used to program the system. Knowledge representation programs control the way information is stored and defined. The goal is to develop a knowledge representation language which can maximize the efficiency of a computers ability to deal with real world problems.

Problem Solving

Computers of today only know what we tell them. Therefore their responses are dependent on the knowledge that we give them. Most times computers derive the correct solutions or better solutions if they are given the correct set of rules. But if the rules are incomplete or the data is incorrect the solution is most likely to be incorrect.

In order for a computer to derive a solution is must first know which type of problem it is being faced with. Artificial Intelligence scientist have created four types of problem categories, single state problems, multiple state problems, contingency problems and exploration problems.

Single State Problems

Multiple State Problems

Contingency Problems

Exploration Problems

After a system has determined the type of problem it is handling it must search for the resolution to the problem. Several search techniques have been developed to determine the best solutions for resolving problems which may be asked of computer systems.

Searching Techniques

Search Tree

Breadth - First Search

Uniformed Cost Search

Depth First Search

Depth Limited Search

Iterative Deepening Search

Bi-directional Search

Reasoning

Store and Fetch

Frame Systems or Semantic Networks

Descriptive Logic Systems

Table Based Indexing

Tree Based Indexing

Once the data storage method has been chosen the reasoning method needs to be determined. The main reasoning methods are theorem prover, productions systems, frame systems, semantic network systems, and descriptive logic systems.

Theorem Prover

• Facts about the problem.
• Background knowledge of the problem.
• A set of rules and equations which assist in updating the system.
• A heuristic function which states the rules of how a system should process and filter.

The theorem prover takes each fact about the problem and matches it up to the background data it has about the problem to come to a resolution. It will then update its data accordingly.

Production Systems

• Working memory which is where the new data is stored to be manipulated
• Rules describing the acceptable actions for the system to take.
• A match phase in which the system matches stored rules with the new data.

The production systems review the suggestions from the match phase and determine if they are correct by using conflict resolution. Some conflict resolution rules would be:

• Has this new rule already been performed, if so do not repeat.
• Rules that state the new rules should only be processed on new data.
• Perform the new rules only on data containing certain variables

Frame Systems and Semantic Networks

These systems concentrate on categories and the links between them. An example of a category would be fish with the item of trout. Frame systems solve problems by determining which category the problem falls into. Then it uses the information about that category to resolve the problem.

Descriptive Logic Systems

Descriptive Logic Systems work much like Frame Systems and Semantic Networks except the categories are defined with greater detail. The problem solving technique is the same as that for Frame Systems and Semantic Networks but the answers tend to be more accurate or descriptive due to the nature of the file being more descriptive.

Pattern Matching / Mixed Parts

Pattern matching and mixed parts problems are instances when the computer system must distinguish an item from its surroundings such as a winter hat in a closet full of clothes. Currently computers have great difficulty resolving pattern matching or mixed parts equations. The techniques have not been developed to allow computers the ability to be able to distinguish between such things and dogs and cats or to pick out an object from a pile of objects. So far this is one ability that escapes computer capability.

Adapting to New Environments

Adapting to new environments is another difficult process to teach a computer. It is impossible to program all the possible environments that a computer may find itself in. In order to teach a computer to adapt to a new environment we must first teach it to learn.

Machine Learning

Most learning systems have four separate components, critic, the learning element, problem generator, and performance element.

Critic

Critic is a function which informs the learning element if the agent is succeeding. This functions is used to ensure the system knows it is performing correctly. If a system were not advised that is was performing correctly it may try to alter its solution path causing a more costly or ineffective path to be taken.

Learning Element

The Learning Element is responsible for making improvements in the systems actions by learning from its mistakes.

Problem Generator

The Problem Generator is responsible for making suggestions about possible actions. This function often causes the system to explore other avenues which may or may not uncover better solutions paths than that stored in memory.

Performance Element

The performance element is responsible for carrying out the action which was decided on.

Learning systems work much like learning by trial and error. They will explore certain avenues then store the result which has the least cost. Much like we do during life we may try several solutions to a problem, we are more likely to remember the solution which caused us the least distress.

Computer Vision

• Cameras which contain light sensitive chips.
• Electromagnetic spectrum
• Infrared sight
• X-ray vision

Most computers to date are using 3D vision. 3D vision is accomplished by three elements, segmentation, orientation of the object and determining the shape of the object.

Segmentation

Determining the Position and Distance of an Object

Determining Shape

A lot more goes into our sight then just seeing the objects around us. The image that we see is passed onto our brain which works with past experiences to determine what we are actually seeing. The simplest things such as walking through a room require much more than just seeing what is around us. Exactly what do we go through when we walk through a room.

1. Our brain takes in the visual display of the room and determines if anything is obstructing our path.
2. If there is an obstacle in our path our brain makes the determination of exactly how far the object is from us and determines what it is and if it needs to be avoided.
3. If the object needs to be avoided our brain determines at which moment we need to start avoiding the object.
4. Once we have avoided the object our brain determines the alternate route too successfully reach our destination.

Computers must first be given some background knowledge of certain obstacles they may encounter such as which can be stepped on, easily moved or avoided. Computers can also posses sensors which assist them in problems such as walking through a room. Such sensors include:

• Acoustical Arrays for hearing
• Smoke and gas detectors for smell
• Force and Torque sensors for touch

Even with all this knowledge the task of walking through a crowded room is still a difficult task for a computer.

Robotics

There are several body parts which have been created to assist computers in their duties, although many of these do no resemble human parts they do the job.

Arms

Hands

Robotic hands are called grippers and are designed to grasp and hold. Robotic hands may consist of:

• A thumb and first finger
• Magnets
• Suctions Cups
• Specific tools such as spray guns, welders or grinders.

Legs and Feet

Many robots have wheels instead or legs of feet. Other robots move on tracks or treads.

Exactly how close are we to passing the Turing Test. Currently computers surpass human intelligence when the problems given to them are logical and the rules to solve the problems are consistent. Computers still have great difficulty with task such as speech, learning and uncertain reasoning which do not fall into that situation. Once computers have all the required capabilities the big problem will be to get all this intelligence into one system. If computer technology continues to advance at its current rate many of us may see the first computer to pass "The Turing Test" in our life times.

Back To My Home Page