Goals Artificial intelligence

1 goals

1.1 reasoning, problem solving
1.2 knowledge representation
1.3 planning
1.4 learning
1.5 natural language processing
1.6 perception
1.7 motion , manipulation
1.8 social intelligence
1.9 creativity
1.10 general intelligence

goals

the overall research goal of artificial intelligence create technology allows computers , machines function in intelligent manner. general problem of simulating (or creating) intelligence has been broken down sub-problems. these consist of particular traits or capabilities researchers expect intelligent system display. traits described below have received attention.

reasoning, problem solving

early researchers developed algorithms imitated step-by-step reasoning humans use when solve puzzles or make logical deductions. late 1980s , 1990s, ai research had developed methods dealing uncertain or incomplete information, employing concepts probability , economics.

for difficult problems, algorithms can require enormous computational resources—most experience combinatorial explosion : amount of memory or computer time required becomes astronomical problems of size. search more efficient problem-solving algorithms high priority.

human beings ordinarily use fast, intuitive judgments rather step-by-step deduction ai research able model. ai has progressed using sub-symbolic problem solving: embodied agent approaches emphasize importance of sensorimotor skills higher reasoning; neural net research attempts simulate structures inside brain give rise skill; statistical approaches ai mimic human ability guess.

knowledge representation

an ontology represents knowledge set of concepts within domain , relationships between concepts.

knowledge representation , knowledge engineering central ai research. many of problems machines expected solve require extensive knowledge world. among things ai needs represent are: objects, properties, categories , relations between objects; situations, events, states , time; causes , effects; knowledge knowledge (what know other people know); , many other, less researched domains. representation of exists ontology: set of objects, relations, concepts, , properties formally described software agents can interpret them. semantics of these captured description logic concepts, roles, , individuals, , typically implemented classes, properties, , individuals in web ontology language. general ontologies called upper ontologies, attempt provide foundation other knowledge acting mediators between domain ontologies cover specific knowledge particular knowledge domain (field of interest or area of concern). such formal knowledge representations suitable content-based indexing , retrieval, scene interpretation, clinical decision support, knowledge discovery via automated reasoning (inferring new statements based on explicitly stated knowledge), etc. video events represented swrl rules, can used, among others, automatically generate subtitles constrained videos.

among difficult problems in knowledge representation are:

default reasoning , qualification problem
many of things people know take form of working assumptions . example, if bird comes in conversation, people typically picture animal fist sized, sings, , flies. none of these things true birds. john mccarthy identified problem in 1969 qualification problem: commonsense rule ai researchers care represent, there tend huge number of exceptions. nothing true or false in way abstract logic requires. ai research has explored number of solutions problem.
the breadth of commonsense knowledge
the number of atomic facts average person knows large. research projects attempt build complete knowledge base of commonsense knowledge (e.g., cyc) require enormous amounts of laborious ontological engineering—they must built, hand, 1 complicated concept @ time. major goal have computer understand enough concepts able learn reading sources internet, , able add own ontology.
the subsymbolic form of commonsense knowledge
much of people know not represented facts or statements express verbally. example, chess master avoid particular chess position because feels exposed or art critic can take 1 @ statue , realize fake. these non-conscious , sub-symbolic intuitions or tendencies in human brain. knowledge informs, supports , provides context symbolic, conscious knowledge. related problem of sub-symbolic reasoning, hoped situated ai, computational intelligence, or statistical ai provide ways represent kind of knowledge.

planning

a hierarchical control system form of control system in set of devices , governing software arranged in hierarchy.

intelligent agents must able set goals , achieve them. need way visualize future—a representation of state of world , able make predictions how actions change it—and able make choices maximize utility (or value ) of available choices.

in classical planning problems, agent can assume system acting in world, allowing agent of consequences of actions. however, if agent not actor, requires agent can reason under uncertainty. calls agent can not assess environment , make predictions, evaluate predictions , adapt based on assessment.

multi-agent planning uses cooperation , competition of many agents achieve given goal. emergent behavior such used evolutionary algorithms , swarm intelligence.

learning

machine learning, fundamental concept of ai research since field s inception, study of computer algorithms improve automatically through experience.

unsupervised learning ability find patterns in stream of input. supervised learning includes both classification , numerical regression. classification used determine category belongs in, after seeing number of examples of things several categories. regression attempt produce function describes relationship between inputs , outputs , predicts how outputs should change inputs change. in reinforcement learning agent rewarded responses , punished bad ones. agent uses sequence of rewards , punishments form strategy operating in problem space. these 3 types of learning can analyzed in terms of decision theory, using concepts utility. mathematical analysis of machine learning algorithms , performance branch of theoretical computer science known computational learning theory.

within developmental robotics, developmental learning approaches elaborated upon allow robots accumulate repertoires of novel skills through autonomous self-exploration, social interaction human teachers, , use of guidance mechanisms (active learning, maturation, motor synergies, etc.).

natural language processing

a parse tree represents syntactic structure of sentence according formal grammar.

natural language processing gives machines ability read , understand human language. sufficiently powerful natural language processing system enable natural language user interfaces , acquisition of knowledge directly human-written sources, such newswire texts. straightforward applications of natural language processing include information retrieval, text mining, question answering , machine translation.

a common method of processing , extracting meaning natural language through semantic indexing. although these indexes require large volume of user input, expected increases in processor speeds , decreases in data storage costs result in greater efficiency.

perception

machine perception ability use input sensors (such cameras, microphones, tactile sensors, sonar , others) deduce aspects of world. computer vision ability analyze visual input. few selected subproblems speech recognition, facial recognition , object recognition.

motion , manipulation

the field of robotics closely related ai. intelligence required robots handle tasks such object manipulation , navigation, sub-problems such localization, mapping, , motion planning. these systems require agent able to: spatially cognizant of surroundings, learn , build map of environment, figure out how 1 point in space another, , execute movement (which involves compliant motion, process movement requires maintaining physical contact object).

social intelligence


kismet, robot rudimentary social skills

affective computing study , development of systems can recognize, interpret, process, , simulate human affects. interdisciplinary field spanning computer sciences, psychology, , cognitive science. while origins of field may traced far philosophical inquiries emotion, more modern branch of computer science originated rosalind picard s 1995 paper on affective computing . motivation research ability simulate empathy, machine able interpret human emotions , adapts behavior give appropriate response emotions.

emotion , social skills important intelligent agent 2 reasons. first, being able predict actions of others understanding motives , emotional states allow agent make better decisions. concepts such game theory, decision theory, necessitate agent able detect , model human emotions. second, in effort facilitate human–computer interaction, intelligent machine may want display emotions (even if not experience emotions itself) appear more sensitive emotional dynamics of human interaction.

creativity

a sub-field of ai addresses creativity both theoretically (the philosophical psychological perspective) , practically (the specific implementation of systems generate novel , useful outputs).

general intelligence

many researchers think work incorporated machine artificial general intelligence, combining skills mentioned above , exceeding human ability in or these areas. few believe anthropomorphic features artificial consciousness or artificial brain may required such project.

many of problems above require general intelligence solved. example, specific straightforward tasks, machine translation, require machine read , write in both languages (nlp), follow author s argument (reason), know being talked (knowledge), , faithfully reproduce author s original intent (social intelligence). problem machine translation considered ai-complete , of these problems need solved simultaneously in order reach human-level machine performance.