Expert Systems | Research & Encyclopedia Articles

Expert Systems

One of the major areas of interest in artificial intelligence is the building of expert systems. An expert system contains information, and searches for patterns in problems it is given to find solutions to them. It is called an expert system because it attempts to closely model the manner in which a human expert would seek a solution when faced with the same problem. A human expert would analyze the situation by comparing it with other known and solved cases, or by applying rules or heuristics, in order to reach a conclusion. The expert system attempts to emulate this behavior. Expert systems can be of several kinds: rule-based, fuzzy logic, neural networks, Bayesian networks, etc.

In order to function, an expert system needs a knowledge base consisting of the facts, expertise, rules of thumb, and any other such that a human expert in the domain would apply in order to arrive at a solution to a problem. A knowledge base may encompass the learning of several experts, or else may be derived from the learning of a single one. It may consist of representative objects within a sub-class. Commonly, knowledge bases focus on narrow issues within a fact situation. They are derived by drawing up case scenarios and having human experts describe their responses to them. These responses are then coded in a certain format to derive the knowledge base.

Rule-based reasoning in expert systems is when an inference system of the form if-then-else is applied. The rules to be applied are simply patterns, and the expert system searches through the patterns until it finds a rule that applies to the given facts. If a condition is true, then something is to be done, else something different is to be done, etc. Rules can be forward-chaining, wherein a set of facts is given and a search is made for any rules that might apply. This sort of reasoning is also called data-driven reasoning. The other kind of reasoning is goal-driven reasoning, also called backward-chaining, wherein a goal is first described, and the expert system tries to come up with facts that can lead to that goal.

Fuzzy logic is another technique in expert systems where instead of the bimodal values TRUE and FALSE (which may be represented by the values 1 and 0), we have a continuum of truth values in the range from 0 to 1, to cover notions such as "probably true," or "mostly false." The definitions of the Boolean operators AND, OR, NOT, etc., and the if-then-else, rules, etc., are all extended to the continuous domain. Fuzzy logic avoids the problems inherent in being forced to make clear-cut choices, and often allows for the representation of guesses, intuitions, and vague reasoning that an expert may make.

A neural network is another kind of computing paradigm, implemented either in specialized hardware or in software, or in some combination of the two. The purpose of neural networks was to emulate the performance of the human brain by copying its physical structure. Almost all neural networks have some manner in which the network can be "trained," and a neural network can learn from examples, just as a child learns to recognize dogs from seeing examples of dogs. This structural capacity for generalization from being introduced to instances can be exploited by interaction with a subject expert, to form an expert system.

Bayesian networks, also called belief networks or probabilistic networks, are systems that use the principles of probability to reason under conditions of uncertainty. Their main advantage over neural networks and other systems, is that they can "explain" their reasoning in arriving at the conclusions they offer. Bayesian networks have been used in statistics, artificial intelligence, medical diagnosis, and other areas.

Expert systems are well suited in cases where a human expert wishes to use a system that performs an independent study of the same facts and offers a "second opinion" as an aid to clarify or verify the human's analysis. They are also suited to cases where a human wishes to use an advanced aid to find relevant facts, rules, and the like that apply to a given situation. However, humans are known to continually apply new paradigms, learn new concepts or move away from old ones, and otherwise refine or enhance their knowledge, but it is not always trivial to teach an old expert system new tricks. A human is also often able to modulate her response, so that a physician may make a quick diagnosis when time is not available and the problem is not critical, while making a detailed diagnosis otherwise. It is not easy to get an expert system to make scaled responses to similar situations. Human experts also often disagree among themselves as to the best course of action, and a single human expert may also pursue or suggest different courses of actions at different types based on subtleties of fact, intuition, or idiosyncrasy. These variations are not easy to incorporate in an expert system.

Expert systems can also not be used in place of humans in most applications because of legal and social concerns. Many applications such as the practice of law or medicine that an expert system could be used for require special licenses from appropriate authorities before someone is considered qualified as a subject expert. Human society of the present day has not evolved to the point where a computerized expert system could be certified as a medical doctor and carry out a practice, and computer scientists agree that expert systems are not nearly as comprehensive as their human counterparts. There is also the question of legal liability--a human expert can be held responsible for any incorrect analyses or actions that cause harm to others, but what about expert systems? Until this question is answered, expert systems will continue to be merely a tool to aid humans. In most cases, they are also not used by lay humans, but only by experts as specialized aids.