Upload
ashlee-wiggins
View
223
Download
6
Embed Size (px)
Citation preview
처음 페이지로 이동
Chapter 11: Analytical Learning
Inductive learning
training examples Analytical learning
prior knowledge + deductive reasoning Explanation based learning - prior knowledge : analyze, explain how each training
example satisfies the target concept
- distinguish relevant features
= generalization based on logical reasoning
- applied to learning search control rules
처음 페이지로 이동
Introduction
Inductive learning
poor performance when insufficient data
Explanation based learning
(1) accept explicit prior knowledge
(2) generalize more accurately than inductive system
(3) prior knowledge = reduce complexity of hypotheses space
reduce sample complexity
improve generalization accuracy
처음 페이지로 이동
Task of learning play chess
- target concept chessboard positions in which black will lose its queen within two moves - human explain/analyze training examples by prior knowledge - knowledge = legal rules of chess
처음 페이지로 이동
Chapter summary
- Learning algorithm that automatically construct/learn from explanation
- Analytical learning problem definition
- PROLOG-EBG algorithm
- General properties, relationship to inductive learning algorithm
- Application to improving performance at large state-space search problems
처음 페이지로 이동
Inductive Generalization Problem
Given: Instances
Hypotheses
Target Concept
Training examples of target concept
Determine:
Hypotheses consistent with the training examples
처음 페이지로 이동
Analytical Generalization Problem
Given: Instances
Hypotheses
Target Concept
Training Examples
Domain Theory
Determine: Hypotheses consistent with training examples and
domain theory
처음 페이지로 이동
Example of an analytical learning problem
Instance space : describe a pair of objects
Color, Volume, Owner, Material, Density, On
Hypothesis space H
SafeToStack(x,y) Volume(x,vx) ^ Volume(y,vy) ^ LessThan(vx,vy)
Target concept : SafeToStack(x,y)
“pairs of physical objects, such that one can be stacked safely on the other”
처음 페이지로 이동
Training Examples On(Obj1, Obj2) Owner(Obj1, Fred) Type(Obj1, Box) Owner(Obj2,Louise) Type(Obj2, Endtable) Density(Obj1,0.3) Color(Obj1,Red) Material(Obj1,Cardboard) Color(Obj2,Blue) Material(Obj2,Wood) Volume(Obj1,2)
Domain Theory B SafeToStack(x,y) ~Fragile(y) SafeToStack(x,y) Lighter(x,y) Lighter(x,y) Weight(x,wx) ^ Weight(y,wy) ^ LessThan(wx,wy) Weight(x,w) Volume(x,v) ^ Density(x,d) ^ Equal(w,times(v,d)) Weight(x,5) Type(x,Endtable) Fragile(x) Material(x,Glass) .......
처음 페이지로 이동
Domain Theory B
- explain why certain pairs of objects can be safely stacked on one another
- described by a collection of Horn clause : enable system to incorporate any learned hypotheses into subsequent domain theories
처음 페이지로 이동
Learning with Perfect Domain Theories : PROLOG-EBG
Correct = assertions are truthful statements Complete = covers every positive examples
Perfect domain theory is available? Yes Why does it need to learn when perfect domain theory
is given?
처음 페이지로 이동
PROLOG-EBG
Operation (1) Leaning a single Horn clause rule
(2) Removing positive examples covered
(3) Iterating this process
Given a complete/correct domain theory
--> output a hypothesis (correct, cover observed
positive training examples)
처음 페이지로 이동
PROLOG-EBG AlgorithmPROLOG-EBG(Target Concept,Training Examples,Domain Theory)• Learned Rules {}• Pos the positive examples from Training Examples• for each Positive Examples in Pos that is not covered by Learned Rules, do
1. Explain:
- Explanation explanation in terms of Domain Theory that Positive
Examples satisfies the Target Concept
2. Analyze:
- Sufficient Conditions the most general set of features of Positive
Examples sufficient to satisfy the Target Concept according to the
Explanation
3. Refine:
- Learned Rules Learned Rules + NewHornClause
Target Concept Sufficient Conditions• Return Learned Rules
처음 페이지로 이동
처음 페이지로 이동
Weakest Preimage
The weakest preimage of a conclusion C with respect to a proof P is the most general set of initial assertions A, such that A entails C according to P
the most general rules
SafeToStack(x,y) Volume(x,vx) ^ Density(x,dx) ^ Equal(wx,times(vx,dx)) ^ LessThan(wx,5) ^
Type(y,Endtable)
처음 페이지로 이동
처음 페이지로 이동
처음 페이지로 이동
Remarks on Explanation-Based Learning
Properties
(1) justified general hypotheses by using prior knowledge
(2) explanation determines relevant attributes
(3) regressing the target concept allows deriving more general
constraints
(4) learned Horn clause = sufficient condition to satisfy target
concept
(5) implicitly assume the complete/correct domain theory
(6) generality of the Horn clause depends on the formulation of
the domain theory
처음 페이지로 이동
Perspectives on example based learning
(1) EBL as theory-guided generalization (2) EBL as example-guided reformation of theories (3) EBL as “just” restating what the learner already “knows”
Knowledge compilation
- reformulate the domain theory to produce general rules that classify examples in a single inference step - transformation = efficiency improving task without altering
correctness of system’s knowledge
처음 페이지로 이동
Characteristics Discovering New Features
- learned feature : feature by hidden units of neural networks
Deductive Learning - background knowledge of ILP : enlarge the set of hypotheses
- domain theory : reduce the set of acceptable hypotheses
Inductive Bias - inductive bias of PROLOG-EBG = domain theory B
- Approximate inductive bias of PROLOG-EBG
= domain theory B +
preference for small sets of maximally general Horn clauses
처음 페이지로 이동
LEMMA-ENUMERATOR algorithm - enumerate all proof trees - for each proof tree, calculate the weakest preimage and construct a Horn clause - ignore the training data - output a superset of Horn clauses output by PROLOG-EBG
Role of training data focus algorithm on generating rules that cover the distribution of instances that occur in practice
Observed positive example allow generalizing deductively to other unseen instances IF ((PlayTennis = YES) (Humidity=x)) THEN ((PlayTennis = YES) (Humidity <= x)
처음 페이지로 이동
Knowledge-level learning - the learned hypothesis entails predictions that go beyond those entailed by the domain theory - deductive closure : set of all predictions entailed by a set of
assertions
Determinations - some attribute of the instance is fully determined by certain other attributes, without specifying the exact nature of the dependency - example target concept : “people who speak Portuguese” domain theory : “ the language spoken by a person is determined
by their nationality” training example : “Joe, 23-year-old Brazilian, speaks Portuguese” conclusion : “all Brazilians speak Portuguese”
처음 페이지로 이동
Explanation-based Learning of Search Control Knowledge
Speed up complex search programs
Complete/Correct domain theory for learning search control knowledge
= definition of legal search operator
+ definition of the search objective
Problem find a sequence of operators that will transform an arbitrary initial
state S to some final state F that satisfies the goal predicate G
처음 페이지로 이동
PRODIGY
Domain-independent planning system
find a sequence of operators that leads from S to O
means-ends planner decompose problems into subgoals
solve them
combine their solution into a solution for the full problem
처음 페이지로 이동
SORA System
Support a broad variety of problem-solving strategies
Learned by explaining situations in which its current strategy leads to inefficiencies
처음 페이지로 이동
Practical Problems applying EBL to learning search control
The number of control rules that must be learned is very large
(1) efficient algorithms for matching rules
(2) utility analysis : estimating the computational cost and
benefit of each rule
(3) identify types of rules that will be costly to match
re-expressing such rules in more efficient forms
optimizing rule-matching algorithm
처음 페이지로 이동
Constructing the explanations for the desired target concept is intractable
(1) example - states for which operator A leads toward the optimal solution
(2) “lazy” or “incremental” explanation - heuristics are used to produce partial/approximate and tractable explanation - learned rules may be imperfect - monitoring performance of the rule on subsequent cases - when error, original explanation is elaborated to cover new case, - more refined rules is extracted