200 likes | 400 Views
Knowledge in Learning. System Architecture LAB Heonkyu Park ’ 00. 5. 31. Where are we?. Focus on Machine Learning Learning from Observation (Ch 18) Learning in Neural Network (Ch 19) Learning of Belief Network (Ch 19) Reinforcement Learning (Ch 20) Knowledge in Learning (Ch 21).
E N D
Knowledge in Learning System Architecture LAB Heonkyu Park ’00. 5. 31 Knowedge in Learning
Where are we? • Focus on Machine Learning • Learning from Observation (Ch 18) • Learning in Neural Network (Ch 19) • Learning of Belief Network (Ch 19) • Reinforcement Learning (Ch 20) • Knowledge in Learning (Ch 21) Knowedge in Learning
Contents • Overview • Knowledge in Learning • Explanation-Based Learning • Learning Using Relevance Information • Inductive Logic Programming • Summary Knowedge in Learning
Overview • So far, constructing a program that has the input/output behavior observed in data. • searching a hypothesis space to find a suitable program. • Study learning methods can take advantage of prior knowledge about the environment. Knowedge in Learning
Knowledge in Learning • Entailment constraint : Hypothesis ∧Descriptions = Classifications • Cumulative Learning Knowedge in Learning
Some Simple Examples First meeting with a native speaker Toasting a sausages Observing consulting session Measuring conductance of a material Knowedge in Learning
Some general Schema • Explanation-based Learning(EBL) Hypothesis ∧ Descriptions = Classifications Background = Hypothesis • Relevance-based Learning(RBL) Hypothesis ∧ Descriptions = Classifications Background ∧ Descriptions ∧ Classifications = Hypothesis • Knowledge-based Inductive L.(KBIL) Background ∧Hypothesis ∧Descriptions = Classifications • Inductive Logic Programming(ILP) Knowedge in Learning
Explanation Based Learning • Basic Idea • Construct an explanation of the observation using prior knowledge • Establish a definition of the class of cases • Example rules Rewrite(u,v) ∧Simplify(v,w) Simplify(u,w) Primitive(u) Simplify(u,u) ArithmeticUnknown(u) Primitive(u) Rewrite( 1 x u, u) Rewrite( 0 + u, u) ……… Knowedge in Learning
EBL (cont’d) • Proof trees for the simplification problem Knowedge in Learning
EBL (cont’d) • Basic EBL process works : • Given example, construct a proof that goal predicate applies to example using available background knowledge • In parallel, construct a generalized proof tree for variabilized goal using same inference steps • Construct a new rule, left-hand side consists of the leaves of the proof tree, right-hand side is variabilized goal • Drop any conditions that are true regardless of the values of variables in the goal Knowedge in Learning
Using Relevance Information • Example : traveler in Brazil • x,y,n,l Nationality(x,n)∧ Nationality(y,n)∧ Language(x,l) • Language(y,l) Nationality(Fernando,Brazil)∧ Language(Fernando,Portuguese) • So, x Nationality(x,Brazil) Language(x,Portuguese) • Functional Dependencies(Determinations) • Nationality(x,n)Language(x,l) • Material(x,m) ∧ Temperature(x,t) Conductance(x,) Material(x,m) ∧ Temperature(x,t) Density(x,) Knowedge in Learning
RBL (cont’d) • Algorithm for finding a minimal consistent determination Knowedge in Learning
Inductive Logic Programming • Entailment Constraint Background ∧Hypothesis ∧Descriptions = Classifications • Example in ‘Family Tree’ • Term : Mother(), Father(), Married() • Properties : Male(), Female() • Target : Grandparent, BrotherInLaw, Ancestor… • Description Father(아빠, 나), Father(나, 아들) Mother(엄마, 남동생), Mother(아내, 딸) Married(아빠, 엄마), Married(누나, 자형) Male(아빠), Male(자형), Female(처제), Female(이모) ……………………………… Knowedge in Learning
ILP (Cont’d) • Classification Grandparent(할머니, 나) Grandparent(아빠,아들) Grandparent(이모부, 딸), Grandparent(할머니,처제) • Hypothesis Grandparent(x, y) [z Mother(x,z) ∧ Mother(z,y)] V [z Mother(x,z) ∧ Father(z,y)] V [z Father(x,z) ∧ Mother(z,y)] V [z Father(x,z) ∧ Father(z,y)] Knowedge in Learning
ILP (Cont’d) • Inverse resolution • Generating inverse proofs • Individual backward steps • Resolvent C resolves into C1 and C2 • Discovering new predicates & knowledge • Top-down learning methods • Start with very general rule • Gradually specialize it Knowedge in Learning
Summary of this chapter • The use of prior knowledge leads to a picture of cumulative learning • Entailment constraints helps to define a variety of learning techniques • EBL extracts general rules from single examples by explaining the examples and generalizing the explanation • RBL uses prior knowledge to identify the relevant attributes • KBIL finds inductive hypotheses that explain sets of observations with background knowledge • ILP techniques perform KBIL using knowledge expressed in first-order logic Knowedge in Learning
SUMMARY of this course • AI (Intro, Intelligent Agents) • Problem Solving • Problem solving by searching, CSP(BFS,DFS,IDS,BS) • Informed Search Method(BFS,IDA*,IIA,GA) • Artificial Life, Game playing • Uncertain Knowledge and Reasoning • Uncertainty, Probability, Bayes’ rule • Probabilistic Reasoning System(Belief Network) • Making Simple/Complex Decision(Utility Fn, Decision Network) • Learning • From observation, Neural Network, Belief Network, Reinforcement, Knowledge in Learning Knowedge in Learning
Thanks you. Knowedge in Learning