Encyclopedia of Computer Science and Technology

artificial intelligence 27to the “goal state.” A properly implemented production systemcannot only solve problems, it can give an explanationof its reasoning in the form of a chain of rules that wereapplied.The program SHRDLU, developed by Marvin Minsky’steam at MIT, demonstrated that within a simplified “microworld”of geometric shapes a program can solve problemsand learn new facts about the world. Minsky later developeda more generalized approach called “frames” to provide thecomputer with an organized database of knowledge aboutthe world comparable to that which a human child assimilatesthrough daily life. Thus, a program with the appropriateframes can act as though it understands a story abouttwo people in a restaurant because it “knows” basic factssuch as that people go to a restaurant to eat, the meal iscooked for them, someone pays for the meal, and so on.While promising, the frames approach seemed to founderbecause of the sheer number of facts and relationshipsneeded for a comprehensive understanding of the world.During the 1970s and 1980s, however, expert systems weredeveloped that could carry out complex tasks such as determiningthe appropriate treatment for infections (MYCIN)and analysis of molecules (DENDRAL). Expert systemscombined rules of inference with specialized databases offacts and relationships. Expert systems have thus been ableto encapsulate the knowledge of human experts and make itavailable in the field (see expert systems and knowledgerepresentation).The most elaborate version of the frames approach hasbeen a project called Cyc (short for “encyclopedia”), developedby Douglas Lenat. This project is now in its thirddecade and has codified millions of assertions about theworld, grouping them into semantic networks that representdozens of broad areas of human knowledge. If successful,the Cyc database could be applied in many differentdomains, including such applications as automatic analysisand summary of news stories.Bottom-Up ApproachesSeveral “bottom-up” approaches to AI were developed inan attempt to create machines that could learn in a morehumanlike way. The one that has gained the most practicalsuccess is the neural network, which attempts toemulate the operation of the neurons in the human brain.Researchers believe that in the human brain perceptions orthe acquisition of knowledge leads to the reinforcement ofparticular neurons and neural paths, improving the brain’sability to perform tasks. In the artificial neural network alarge number of independent processors attempt to performa task. Those that succeed are reinforced or “weighted,”while those that fail may be negatively weighted. This leadsto a gradual improvement in the overall ability of the systemto perform a task such as sorting numbers or recognizingpatterns (see neural network).Since the 1950s, some researchers have suggested thatcomputer programs or robots be designed to interact withtheir environment and learn from it in the way that humaninfants do. Rodney Brooks and Cynthia Breazeal at MIThave created robots with a layered architecture that includesmotor, sensory, representational, and decision-making elements.Each level reacts to its inputs and sends informationto the next higher level. The robot Cog and its descendantKismet often behaved in unexpected ways, generating complexresponses that are emergent rather than specificallyprogrammed.The approach characterized as “artificial life” adds agenetic component in which the successful componentspass on program code “genes” to their offspring. Thus, thepower of evolution through natural selection is simulated,leading to the emergence of more effective systems (seeartificial life and genetic algorithms).In general the top-down approaches have been moresuccessful in performing specialized tasks, but the bottomupapproaches may have greater general application, as wellas leading to cross-fertilization between the fields of artificialintelligence, cognitive psychology, and research intohuman brain function.Application AreasWhile powerful artificial intelligence is not yet ubiquitousin everyday computing, AI principles are being successfullyused in a number of application areas. These areas, whichare all covered separately in this book, include• devising ways of capturing and representing knowledge,making it accessible to systems for diagnosis andanalysis in fields such as medicine and chemistry (seeknowledge representation and expert systems)• creating systems that can converse in ordinary languagefor querying databases, responding to customerservice calls, or other routine interactions (see naturallanguage processing)• enabling robots to not only see but also “understand”objects in a scene and their relationships (see computervision and robotics)• improving systems for voice and face recognition, aswell as sophisticated data mining and analysis (seespeech recognition and synthesis, biometrics,and data mining)• developing software that can operate autonomously,carrying out assignments such as searching for andevaluating competing offerings of merchandise (seesoftware agent)ProspectsThe field of AI has been characterized by successive wavesof interest in various approaches, and ambitious projectshave often failed. However, expert systems and, to a lesserextent, neural networks have become the basis for viableproducts. Robotics and computer vision offer a significantpotential payoff in industrial and military applications. Thecreation of software agents to help users navigate the complexityof the Internet is now of great commercial interest.The growth of AI has turned out to be a steeper and morecomplex path than originally anticipated. One view suggestssteady progress. Another, shared by science fiction