Artificial Intelligence an Overview.ppt

《Artificial Intelligence an Overview.ppt》由会员分享，可在线阅读，更多相关《Artificial Intelligence an Overview.ppt（30页珍藏版）》请在麦多课文档分享上搜索。

1、Artificial Intelligence an Overview,Dr. B. Arunkumar Coimbatore Institute of Technology Coimbatore, India,Artificial Intelligence an Overview,What is intelligence? Intelligence is hard to describe. More a performance view rather than a structural one. Intelligence is observed in NEW areas. New areas

2、 where the knowledge is still incomplete. Intelligence - ability to work efficiently with Incomplete, Complex patterns,Artificial Intelligence an Overview,Artificial Intelligence - Enabling computers to work efficiently with Incomplete, Complex Patterns. What is the problem?Incomplete, complex patte

3、rns a large, unbounded search space. Searching this is time consuming Non Polynomial time complexity.,Artificial Intelligence an Overview,More details on patterns: Pattern a set of repeating, significant attributes. Complexity of a pattern measured by the number of attributes and the relationships b

4、etween these attributes. The more attributes The more complex The more relationships (inter dependencies) The more complex.,Artificial Intelligence an Overview,A view of the world: Three segments Segment 1 Totally known segment.All knowledge in this segment is known Methods exist for all problems So

5、lutions are method oriented. Underlying patterns can be ignored. Example - Find the square root of a number.,Artificial Intelligence an Overview,A view of the world: Segment 3 - Totally Unknown Hardly anything of topics in this area is known. Human beings are themselves unable to do much here. Examp

6、le - Life on other planets,Artificial Intelligence an Overview,A view of the world: Segment 2 Partially Known. Quite a lot is known about topics in this segment, but not everything. = Incomplete, Ambiguous patterns. Example Diagnosing diseases.,Artificial Intelligence an Overview,Intelligence is req

7、uired to handle problems in Segment 2. Algorithmic approaches cannot work here as an algorithm, by definition is finite, definite, and effective. (Definite is the opposite of ambiguous.) As more knowledge is acquired, topics in Segment 3 move to Segment 2 and topics in Segment 2 move to Segment 1.,A

8、rtificial Intelligence an Overview,Problem that artificial intelligence attempts to handle is “Providing efficient solutions to problems in an ambiguous, incomplete pattern area”. Artificial intelligence itself lies in Segment 2 of the view of the world. Solution - Non-algorithmic approaches.,Artifi

9、cial Intelligence an Overview,Artificial intelligence techniques can be divided into two types: Symbolic computation Non- symbolic computation,Artificial Intelligence an Overview,Symbolic Computation:Symbol: represents a concept, rather than a value.A symbol represents a relationship among two or mo

10、re classes. (class as in Object Oriented Programming Systems.) Symbolic computation represents an extreme in a continuum: Variable (representing numbers), Data Structure (variables of a particular type), Class (representing a collection of related variables and their functions), Symbol (representing

11、 collection of Objects and the relationships between them),Artificial Intelligence an Overview,Symbolic Computation has two branches Heuristic search Adjoining, Segment 1 of the World view. Heuristic A guide, an approximation, a thumb rule. Basically helps in pruning the search tree. Knowledge-based

12、 systems In the world view,between heuristic search and sub-symbolic computation. Knowledge Data is an understood, recognized format, Information is Useful data and Knowledge is Generalized Information. = Concepts, Patterns.,Artificial Intelligence an Overview,Heuristic Search Two types Proceeds fro

13、m Start state to Goal state A* - Data driven. Proceeds from Goal state to Start state AO* - Goal driven. A* - generates a solution path. Uses heuristics to prune the possible set of operators. AO* - generates a solution tree. Creates sub-goals for a particular goal, until the sub-goal is directly ac

14、hievable.,Artificial Intelligence an Overview,Core areas of Heuristic search: Problem representation - by a State space. Each node in the State space represents a complete state of the problem. Operators Change one state to another. Heuristic Evaluation function Evaluates the goodness of each of the

15、 possible next states. (Not a definite evaluation, only an approximation.),Artificial Intelligence an Overview,The Heuristic evaluation function is basically a form of hill climbing: Take the steepest gradient which will be the shortest path to the peak (goal). Problems in Heuristic Search: Local Ma

16、xima A particular point in the search space may be better than all neighboring points, but still, may not be the ultimate goal. This is called a Local Maxima. Solved by making Random Jumps.,Artificial Intelligence an Overview,Knowledge Based Systems: Core Areas of Knowledge Based systems Knowledge B

17、ase Representation Inference Engine User interface Knowledge acquisition module,Artificial Intelligence an Overview,Representation techniques are primarily: production rules sets of if-then rules, similar to production rules used to specify a grammar. Example: If the car does not start check the bat

18、tery, by pressing the horn.,Artificial Intelligence an Overview,Representation techniques are primarily: Semantic Networks Set of Nodes and Links between them. The links represent Relationships between the nodes Example: Nodes Man, Hands, Legs, Walk Relationships Has (between Man and hands and betwe

19、en Man and Legs) and Can (between Man and Walk). A type of Semantic networks is Frames (Slot-filler notation). These encode default (commonly occurring) values (filler) for the attributes in a relation (slot).,Artificial Intelligence an Overview,Inference Engine - Search on the knowledge base leads

20、to Inferences. Knowledge Acquisition module - The knowledge being incomplete will be dynamic. Provision to acquire knowledge is provided by using machine learning strategies.,Artificial Intelligence an Overview,Machine Learning Strategies: Rote learning The system is told the actual knowledge. The s

21、ystems work is to map the knowledge into its internal representation. Learning by being told The system is given paragraphs that convey the knowledge. The system has to glean the knowledge and then store it. Learning by being told and asking questions In addition to strategy 2, the system analyses t

22、he knowledge, finds discrepancies and asks questions to sort out the conflicts. Learning by induction from positive examples - The system is given examples of the concept. It generalizes the examples to arrive at the knowledge,Artificial Intelligence an Overview,Machine Learning Strategies: Learning

23、 by Induction from Positive examples and Negative examples To avoid over generalization, negative examples are given, which are used to specialize the knowledge. Learning by Induction through experimentation - The system generates examples itself by designing experiments on the environment. Learning

24、 by Analogy The system maps the knowledge it has to the new problem, using analogy. Learning by Abduction The system creates new hypotheses and designs experiments to ratify them.,Artificial Intelligence an Overview,Genetic Programming: This field lies at the extreme of Knowledge Based Systems (adjo

25、ining sub-symbolic computation in the World view) They model Human evolution methods. Approach: Create an initial population of entities Each entitys characteristics are represented A fitness function evaluates the entities. The best two of the population are chosen These two are used to generate of

26、fsprings. new population. Process repeats.,Artificial Intelligence an Overview,Offspring generation operators: Reproduction All characteristics of both parents are reproduced in the offspring. Crossover - A subset of characteristics of one parent are linked with the subset of characteristics of the

27、other parent. Mutation The characteristics of one parent are changed randomly to create the offspring. Handles the Local Maxima problem,Artificial Intelligence an Overview,Sub Symbolic Computation (Neurocomputing): Adjoins Segment 3 of the world view. Deals with signal level computation. Required be

28、cause a number of problems do not have explicit knowledge associated with them. Example recognizing people or recognizing handwriting. This area deals with patterns that are more complex than the ones dealt with by symbolic computation.,Artificial Intelligence an Overview,Core areas of Sub-symbolic

29、computation are: Architecture Learning mechanism In sub-symbolic computation all the knowledge is learnt by the system. Neuro-computing attempts to mimic the structure of the human intelligence system, with its neurons and synapses. Neuron receives input from many other neurons. Each input is magnif

30、ied by a multiplication factor. (This multiplication factor represents the degree of interest, effect that the particular input has on the neuron.),Artificial Intelligence an Overview,All the multiplied values are summed up and compared to a threshold value. If the threshold value is less then the n

31、euron fires an output. Knowledge is acquired by learning the correct multiplication values. Learning is done in one of two ways: Supervised learning - Here the desired output for a given input is known. A simple method is Back Propagation network. Here the output is compared with the desired output.

32、 Differences are propagated backwards, to make changes to the multiplication factors.,Artificial Intelligence an Overview,Unsupervised learning Here the desired output is not given to the system. The system uses Clustering to club similar input together. Example Kohonen A third learning technique is

33、 Self-Supervised Learning - Here the results of a previous iteration are used to bias the clustering results in the current iteration. Example Adaptive Resonance Technique.,Artificial Intelligence an Overview,Applications of Artificial Intelligence: Diagnostic: To diagnose diseases MYCIN, INDUCE Des

34、ign: To design the computer network architecture in an university - R1 Education: Intelligent Tutoring Systems BUGGY,Artificial Intelligence an Overview,Recognize Handwriting NEOCOGNITRON Discover theories in Number system AMArtificial Intelligence differs from conventional computer systems in Being

35、 Non-Algorithmic Being the only systems that Discover the solution and then Execute it. (Other computer systems have the solution designed by the programmer and only execute the solution.),Artificial Intelligence an Overview,References: Artificial Intelligence A new synthesis, Neils J Nillson, Morgan Kaufmann publishers, 1998Artificial Intelligence A modern approach Stuart Russel and Keith Norvig, 1998.,