Explanation-Based Learning (EBL) : One Definition

Explanation-Based Learning (EBL)
One definition:
Learning general problem-solving techniques by observing and analyzing human solutions to specific problems.
The EBL Hypothesis

By understanding why an example is a member of a concept, can learn the essential properties of the concept Trade-off the need to collect many examples for the ability to explain single examples (a domain theory)
Learning by Generalizing Explanations

Given
Goal (e.g., some predicate calculus statement) Situation Description (facts) Domain Theory (inference rules) Operationality Criterion
Use problem solver to justify, using the rules, the goal in terms of the facts. Generalize the justification as much as possible. The operationality criterion states which other terms can appear in the generalized result.
Standard Approach to EBL

An Explanation (detailed proof of goal) goal
facts After Learning (go directly from facts to solution): goal
facts
Unification-Based Generalization
An explanation is an inter-connected collection of pieces of knowledge (inference rules, rewrite rules, etc.) These rules are connected using unification, as in Prolog The generalization task is to compute the most general unifier that allows the knowledge pieces to be connected together as generally as possible
The EGGS Algorithm (Mooney, 1986)

bindings = { }
FOR EVERY equality between patterns P and Q in explanation DO bindings = unify(P,Q,bindings) FOR EVERY pattern P DO P = substitute-in-values(P,bindings) Collect leaf nodes and the goal node
Sample EBL Problem

Initial Domain Theory
knows(?x,?y) AND nice-person(?y) -> likes(?x,?y) animate(?z) -> knows(?z,?z) human(?u) -> animate(?u) friendly(?v) -> nice-person(?v) happy(?w) -> nice-person(?w)
Specific Example
Given human(John) AND happy(John) AND male(John), show that likes(John,John)
Explanation to Solve Problem

likes(John,John) knows(John,John) animate(John) human(John) happy(John) nice-person(John)
Explanation Structure
likes(John,John) likes(?x,?y) knows(?x,?y) knows(?z,?z) animate(?z) animate(?u) human(?u) human(John) happy(?w) happy(John) nice-person(?y) nice-person(?w)
Necessary Unifications:
All variables must match ?z
Resulting Rule:
human(?z) AND happy(?z) -> lik es(?z,?z)
Prototypical EBL Architecture

Specific Goal/Problem (Partial) External Solution
Knowledge Base
Problem Solver (Understander)
New General Concept Generalizer
Explanation
Imperfect Theories and EBL

Incomplete Theory Problem
Cannot build explanations of specific problems because of missing knowledge
Intractable Theory Problem

Have enough knowledge, but not enough computer time to build specific explanation
Inconsistent Theory Problem

Can derive inconsistent results from a theory (e.g., because of default rules)
Some Complications
Inconsistencies and Incompleteness may be due to abstractions and assumptions that make a theory tractable.
Inconsistencies may arise from missing knowledge (incompleteness).
e.g., making the closed-world assumption
Issues with Imperfect Theories

Detecting imperfections
broken explanations (missing clause) contradiction detection (proving P and not P) multiple explanations (but expected!) resources exceeded
Correcting imperfections
experimentation - motivated by failure type (explanationbased) make approximations/assumptions - assume something is true
EBL as Operationalization (Speedup Learning)

Assuming a complete problem solver and unlimited time, EBL already knows how to recognize all the concepts it will know. What it learns is how to make its knowledge operational (Mostow).
Is this learning? Isnt 99% of human learning of this type?
Knowledge-Level Learning (Newell, Dietterich)

Knowledge closure
all things that can be inferred from a collection of rules and facts
Pure EBL only learns how to solve faster, not how to solve problems previously insoluble. Inductive learners make inductive leaps and hence can solve more after learning. What about considering resource-limits (e.g., time) on problem solving?
Negative Effects of Speedup Learning

The Utility Problem
Time wasted checking promising rules
rules that almost match waste more time than obviously irrelevant ones
General, broadly-applicable rules mask more
efficient special cases
Defining Utility (Minton)

Utility = (AvgSav * ApplFreq) - AvgMatchCost
where AvgSav - time saved when rule used ApplFreq - probability rule succeeds given its preconditions tested AvgMatchCost - cost of checking rules preconditions
Rules with negative utility are discarded

estimated on training data
Learning for Search-Based Planners

Two options
1) Save composite collections of primitive operators, called MACROPS
explanation turned into rule added to knowledge base
2) Have a domain theory about your problem solver

use explicit declarative representation build explanations about how problems were solved which choices lead to failure, success, etc. learn evaluation functions (prefer pursuing certain operations in certain situations)
Reasons for Control Rules

Improve search efficiency (prevent going down blind alleys) To improve solution quality (dont necessarily want first solution found via depth-first search) To lead problem solver down seemingly unpromising paths
overcome default heuristics designed to keep problem solver from being overly combinatoric
PRODIGY - Learning Control Knowledge (Minton, 1989)

Have domain theory about specific problem AND another about the problem solver itself Choices to be made during problem solving:
which node in current search tree to expand which sub-goal of overall goal to explore relevant operator to apply binding of variables to operators
Control rules can

lead to the choice/rejection of a candidate lead to a partial ordering of candidates (preferences)
SOAR (Rosenbloom, Laird, and Newell, 1986)

Production system that chunks productions via EBL
Production system - forward chaining rule system for problem solving Key Idea: IMPASSES
occur when system cannot decide which rule to apply solution to impasse generalized into new rule
Summary of SOAR
A Production System with three parts: A general-purpose forward search procedure A collection of operator-selection rules that help decide which operator to apply A look-ahead search procedure invoked when at an impasse When the impasse occurs, can learn new rules to add to collection of operator-selection rules
Reasoning by Analogy
Create a description of a situation with a known solution and then use that solution in structurally similar situations Problem: a doctor can use a beam of radiation to destroy a cancer, but at the high amount needed, it will also destroy the healthy tissue in any path it follows Idea: find a similar (some how) situation and use it to create a solution
Reasoning by Analogy Story

Similar story: a general needs to send his troops to a particular city for a battle by a particular time, but there is no road wide enough to accommodate all of his troops in the time remaining (even though there are several roads) Solution: break up the troops into smaller groups and send each group down a different road How to solve the radiation situation??

Explanation-Based Learning (EBL) : One Definition

Uploaded by

Copyright:

Available Formats

Explanation-Based Learning (EBL) : One Definition

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Explanation-Based Learning (EBL) : One Definition

Uploaded by

Copyright:

Available Formats

Explanation-Based Learning (EBL)

The EBL Hypothesis

Learning by Generalizing Explanations

Standard Approach to EBL

facts After Learning (go directly from facts to solution): goal

The EGGS Algorithm (Mooney, 1986)

Sample EBL Problem

Explanation to Solve Problem

Prototypical EBL Architecture

Problem Solver (Understander)

New General Concept Generalizer

Imperfect Theories and EBL

Intractable Theory Problem

Inconsistent Theory Problem

Issues with Imperfect Theories

EBL as Operationalization (Speedup Learning)

Knowledge-Level Learning (Newell, Dietterich)

Negative Effects of Speedup Learning

General, broadly-applicable rules mask more

efficient special cases

Defining Utility (Minton)

Rules with negative utility are discarded

Learning for Search-Based Planners

2) Have a domain theory about your problem solver

Reasons for Control Rules

PRODIGY - Learning Control Knowledge (Minton, 1989)

Control rules can

SOAR (Rosenbloom, Laird, and Newell, 1986)

Reasoning by Analogy Story

You might also like

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.