- Home
- Documents
- Machine Learning Chapter 5. Artificial IntelligenceChapter 52 Learning 1. Rote learning rote( โรท ) n. วิถีทาง, ทางเดิน, วิธีการตามปกติ, (by rote จากความทรงจำ.

prev

next

of 34

Published on

20-Jan-2016View

222Download

0

Transcript

<ul><li><p>Machine Learning Chapter 5 </p></li><li><p>Learning1. Rote learning rote() n. ,,, (by rote ), S. repetition2. Learning by taking advice3. Learning by problem solvingParameter adjustmentMacro-Operators4. Learning from examplesInduction : Winstons learning program p.458Version Spaces : Candidate eliminate algorithmDecision tree5. Explanation-based learning p 4826. Formal learning theory</p></li><li><p>Winstons learning program</p></li><li><p>Winstons learning programConcept : P.459Begin with a structural description of one known instance of the concept. Call the description the concept definition.Examine descriptions of other known instances of the concepts. Generalize the definition to include them.Examine descriptions of near misses of concept, Restrict the definition to exclude these.</p></li><li><p>HOUSE OF 17.2ARCH OF 17.2ARCH OF 17.2</p></li><li><p>Winstons learning program</p></li><li><p>Winstons learning program</p></li><li><p>Winstons learning program</p></li><li><p>Winstons learning program p.458</p><p>Block world concept : Figure 17.2 p. 459Structure description : Figure 17.3 p. 460The comparison of two arches : Figure 17.4 p. 461The arch description after two examples : Figure 17.5 p. 462The arch description after a near miss : Figure 17.6 p. 463use semantic networks to describe block structuresuse matching process to detect similarities and differences between structuresuse isa hierarchy to describe relationship among already known objects</p></li><li><p>Semantic Networkisaisa</p></li><li><p>Semantic Network</p></li><li><p>Version SpacesThe goal : to produce a description that is consistent with all positive examples but no negative examples in the training set.use frame representing concept for car see Figure 17.7 p. 463Features/Slots : { value1, value2,...,valueN }origin : { Japan, USA, Britain }Variables : X1, X2, X3concept space : see Figure 17.11 Concept of Version Spaces p. 466variablestarget conceptall training instance </p></li><li><p>Version Spaces</p></li><li><p>Version Spacesversion space = current hypothesis = subset of concept space = largest collection of descriptions that is consistent with all the training examples seen so far. concept space = G or SG = contain the most general descriptions consistent with the training example seen so far.S = contain the most specific descriptions consistent with training examplespositive example (+) move S to more specificnegative example (-) move G to more specificif G and S sets converge the hypothesis is a single concept description</p></li><li><p>Version SpacesCandidate Eliminate Algorithm p.466-467 algorithm that use to narrow down the version space by remove any descriptions that are inconsistent with set G and set S Car ExampleFigure 17.7 Concept Car : p. 463Figure 17.8 Representation language for car : p. 464Figure 17.9 The concept Japanese car : p. 464Figure 17.10 Partial ordering of concepts : p. 465Figure 17.12 Positive and negative examples of car : p. 467</p></li><li><p>Version Spaces</p></li><li><p>Version Spaces</p></li><li><p>Version Spaces</p></li><li><p>Version Spaces</p></li><li><p>Candidate Eliminate Algorithm</p></li><li><p>Candidate Eliminate Algorithm</p></li><li><p>Version SpacesWe want Japanese economy car From Figure 17.12 Positive and negative examples of car : p. 467[origin = X1, manufacture = X2, color = X3, decade = X4, type = X5]</p><p>GET EX1 (+) G = {(X1, X2, X3, X4, X5)} S = {(Japan,Honda, Blue,1980,Economy}) =Figure 17.12 in EX1GET EX2 (-) G = {(X1, Honda, X3, X4, X5), (X1, X2, Blue, X4, X5) , (X1, X2, X3, 1980, X5), (X1, X2, X3, X4, Economy)}S = {(Japan,Honda, Blue,1980,Economy}) ** the same because (-) exampleGET EX3 (+) check G first, G = {(X1, X2, Blue, X4, X5) ,(X1, X2, X3, X4, Economy)} S = {(Japan,X2, Blue,X4,Economy}) GET EX4 (-) check G first, G = {(Japan, X2, Blue, X4, X5) , (Japan, X2, X3, X4, Economy)} S = {(Japan,X2, Blue,X4,Economy}) ** the same because (-) exampleGET EX5 (+) check G first, G = {(Japan, X2, X3, X4, Economy)} S = {(Japan,X2, X3,X4, Economy})</p></li><li><p>Version SpacesNote : The algorithm is least commitment algorithm : produce as little as possible at each stepProblems 1.) S and G may not converge to a single hypothesis2. ) if there is a noise (inconsistent data) the algorithm will be premature, we may prune the target concept too fast* For example if the data number three given the negative sign (-) instance of positive sign (+) ... no matter how much the data is we can not find the concept....* How to fix this problem is to maintain several G and S sets BUT it is costly and may have the bounded inconsistency problem3.) We can not use OR in the questions ask * For example : Italian sport car or German luxury car</p></li><li><p>Decision TreeID3 Program = to classify a particular input, we start at the top of the tree and answer questions until we reach a leaf, where the classification is stored. See Figure 17.13 Decision tree p. 4701. Choose window = random subset of training examples to train2. Outside window = use to test the decision tree3. Use empirical evidence (iterative method) to build up decision tree4. Building a node = choosing some attribute to divide training instance into subset consider (+) signCan use with OR .... just change (-) sign into (+) sign Problems : noisy input, attribute value may be unknown, may have large decision tree and hard to understand relationship See Figure 17.13 </p></li><li><p>Decision Tree</p></li><li><p>Explanation-Based Learningprovide explanationdepend on domain theory/ domain knowledge</p></li><li><p>Formal Learning TheoryGiven positive and negative examplesproduce algorithm that will classify future examples correctly with probability 1/hComplexity of learning :the error tolerance (h)the number of binary features present in the examples (t)the size of the rule necessary to make the discrimination (f)</p></li><li><p>Formal Learning Theoryif the number of training examples required is polynomial in h,t, and f then the concept is learnable.few training examples are needed learnablewe restrict the learner to the positive examples only.See Figure 12.22 Concept of elephant P. 483elephant = gray, mammal, large </p></li><li><p>Formal Learning Theory</p></li><li><p>Inductioninduction : A method of reasoning by which one infers a generalization from a series of instances. Inductive syllogisms are of the following form:1. These beans are from this bag. (and these beans..., and these beans..., etc.)2. These beans are (all) white. # 3 Therefore, all beans from this bag are white. In a much broader sense, induction can be thought to include various other forms of reasoning including reasoning, inference to cause form symptoms, and confirmation of laws and theories.1emphasis to all BEANS : all instances</p></li><li><p>Deductiondeduction - A method of reasoning by which one infers a conclusion from a set of sentences by employing the axioms and rules of inference for a given logical system. Use the term 'deduction' in a general sense to denote the fact that a conclusion follows necessarily from the premises.Deductive syllogisms in quantificational predicate calculus are of the following form:1. All beans from this bag are white....2. These beans are from this bag. #4 Therefore, these beans are white..... 1emphasis to one BEAN : one instance</p></li><li><p>Abductionabduction -A method of reasoning by which one infers to the ......best explanation..... - A heuristic procedure that reasons inductively from available empirical evidence to the discovery of the probable hypotheses that would best explain its occurrence.Abductive syllogisms are of the following form:#3 All beans from this bag are white#4 These beans are white. emphasis to one BEANS </p></li><li><p> The End</p></li></ul>