Advanced Reasoning: Explanation-Based Learning and Analogy

Prerequisites

Understanding of case-based reasoning
Familiarity with scripts and frames
Knowledge of learning methods
Completed Fundamentals and Learning Methods modules

Learning Goals

After completing this module, you will be able to:

Apply explanation-based learning to transfer knowledge across situations
Use domain models to explain why solutions work
Perform analogical reasoning across similar and cross-domain problems
Implement the analogical reasoning process: retrieval, mapping, transfer, evaluation
Distinguish between semantic, pragmatic, and structural similarity
Design systems that learn and reason by analogy

1. Explanation-Based Learning (EBL)

Core Concept

Explanation-Based Learning learns general principles from single examples by explaining WHY a solution works, then abstracting the explanation to apply to new situations.

Key Difference from Other Learning:

Case-Based Reasoning: Stores and retrieves specific cases
EBL: Explains one case, extracts general principle, applies broadly

The Cup Problem

Problem: How to transport soup from kitchen to table?

Solution 1: Use a ceramic cup

Question: Will this solution work? WHY?

Explanation:

Goal: Transport soup (liquid)

Why ceramic cup works:
1. Cup has concavity (holds liquids)
2. Cup is liftable (can be carried)
3. Cup is stable (won't tip easily)
4. Ceramic material:
   - Doesn't leak
   - Heat resistant
   - Durable

Causal chain:
Concave shape → holds liquid → enables transport
Liftable + Stable → can carry without spilling

Concept Space and Knowledge

Four Levels:

Level 1: Specific Example

"Use this ceramic cup to transport soup"
Very narrow, only works for this exact cup

Level 2: Example-Based (Case-Based Reasoning)

"Use cups similar to this ceramic cup"
Transfer to similar cups, limited generalization

Level 3: Explanation-Based Learning

"Use any container that:
 - Has concavity (holds liquids)
 - Is liftable
 - Is stable
 - Doesn't leak"

Works for: cups, bowls, pots, bottles, jars
Doesn't require: ceramic, specific shape, color

Level 4: First Principles

"Move liquid by any means that prevents spillage"

Most general, requires deep understanding
May need complex reasoning each time

EBL operates at Level 3:

More general than case-based (Level 2)
More specific than first principles (Level 4)
Practical balance: General enough to transfer, specific enough to be useful

Prior Knowledge Requirements

EBL requires domain knowledge:

Domain Theory:
- Objects have properties (shape, material, size)
- Concave shapes can hold liquids
- Gravity pulls liquids downward
- Liftable objects can be carried
- Stable objects resist tipping
- Some materials are leak-proof

Without this knowledge:

Cannot explain WHY cup works
Cannot abstract general principle
Cannot transfer to new situations

Trade-off:

Strength: Powerful generalization from single example
Limitation: Requires rich domain knowledge

Abstraction Process

Step 1: Explain Specific Solution

Specific: Ceramic cup transports soup

Why?
- Cup shape is concave (holds soup)
- Cup size is liftable (person can carry)
- Cup has handle (easy to grip)
- Ceramic material doesn't leak

Step 2: Identify Essential Features

Essential:
- Concave (必须: needed to hold liquid)
- Liftable (必须: needed to transport)
- Leak-proof (必须: prevents spilling)

Non-Essential:
- Ceramic (not necessary: metal, glass, plastic also work)
- Has handle (helpful but not required)
- Specific size (range of sizes work)
- Color (irrelevant)

Step 3: Abstract to General Principle

General Rule:
"To transport liquid, use container that is:
 1. Concave (holds liquid)
 2. Liftable (can be moved)
 3. Leak-proof (contains liquid)"

Applies to: cups, bowls, bottles, pots, jars, buckets

Transfer to New Problems

New Problem: How to transport juice from refrigerator to table?

Apply Learned Principle:

Need:
- Concave container (holds juice)
- Liftable (can carry from fridge)
- Leak-proof (doesn't spill)

Solution options:
- Glass (✓ concave, liftable, leak-proof)
- Bottle (✓ concave, liftable, leak-proof)
- Bowl (✓ concave, liftable, leak-proof)
- Plate (✗ not concave, can't hold liquid well)

Transfer successful! Principle learned from soup/cup applies to juice/glass.

EBL in Practice

Medical Diagnosis:

Case: Patient with fever, cough, fatigue → Flu

Explanation:
- Fever indicates immune response to infection
- Cough indicates respiratory system involvement
- Fatigue indicates systemic viral infection
- Combination matches influenza pattern

Learned Principle:
"Fever + Respiratory symptoms + Fatigue → Likely viral infection"

Transfer:
Apply to new patients with similar symptom combinations

Engineering:

Case: Bridge design with specific steel beams works

Explanation:
- Load distributed across multiple support points
- Material strength exceeds maximum stress
- Design accounts for wind/seismic forces
- Foundations anchored in bedrock

Learned Principle:
"Successful bridge design requires:
 - Load distribution
 - Material strength margin
 - Dynamic force consideration
 - Stable foundation"

Transfer:
Apply to new bridge designs with different materials/spans

2. Analogical Reasoning

What is Analogy?

Analogy is reasoning about a novel problem by mapping it to a familiar problem where the solution is known, then transferring the solution.

Structure:

Source (familiar) → Target (novel)
Known problem    → New problem
Known solution   → Adapted solution

Example:

Source: Heat flow in metal rod (familiar from physics)
Target: Traffic flow on highway (novel problem)

Mapping:
- Heat → Cars
- Temperature → Density
- Thermal conductivity → Road capacity
- Heat source → On-ramp

Transfer:
Apply heat diffusion equations to traffic flow

The Four-Step Process

Step 1: RETRIEVAL

Given new problem → Find similar past case

Similarity metrics:
- Surface similarity (superficial features)
- Structural similarity (relationships, causality)
- Pragmatic similarity (goal-relevance)

Step 2: MAPPING

Align elements between source and target

Create correspondences:
Source Element A ← corresponds to → Target Element A'
Source Element B ← corresponds to → Target Element B'
Source Relation R ← corresponds to → Target Relation R'

Step 3: TRANSFER

Apply source solution to target problem

Transfer:
- Known facts from source → New facts for target
- Known solution from source → Adapted solution for target
- Causal structure from source → Understanding of target

Step 4: EVALUATION

Test transferred solution

Methods:
- Execute in real world
- Simulate
- Expert review
- Formal verification

Outcomes:
- Success → Store as new case
- Failure → Explain why, learn from mistakes
- Partial → Adapt further, iterate

Three Types of Similarity

1. Semantic Similarity (Surface Features)

Coffee cup vs. Tea cup
Similar:
- Both are cups
- Similar shapes
- Similar functions
- Similar materials

Easy to recognize, limited transfer power

2. Pragmatic Similarity (Goal-Relevance)

Coffee cup vs. Travel mug
Similar because:
- Both serve same goal (contain hot beverages)
- Both portable
- Both human-scale

Different surface features but same function

3. Structural Similarity (Relational Structure)

Heat flow vs. Traffic flow
Different domains, but similar structure:
- Both have flow (heat/cars)
- Both have sources and sinks
- Both have resistance/capacity
- Both follow diffusion-like equations

Most powerful for transfer, hardest to recognize

Spectrum of Similarity

Near Transfer ←─────────────────────→ Far Transfer
(Same Domain)                        (Different Domains)

Examples:
│
├─ Cup to mug (very near)
│  Same domain, similar features
│
├─ Cup to bowl (near)
│  Same domain (containers), different shape
│
├─ Cup to bucket (moderate)
│  Same function (hold liquids), different scale/context
│
├─ Restaurant script to Cafeteria script (far)
│  Different domains, shared abstract structure
│
└─ Heat flow to Traffic flow (very far)
   Completely different domains, shared mathematical structure

Near Transfer:

Easy to recognize
Surface features guide retrieval
Straightforward mapping
Direct solution transfer

Far Transfer (Cross-Domain Analogy):

Hard to recognize
Requires structural understanding
Complex mapping
Solution needs significant adaptation
BUT: Most powerful for innovation and insight

Example: Solar System and Atom Analogy

Historical Analogy (Rutherford’s Atomic Model):

Source: Solar System

- Sun at center (massive, stationary)
- Planets orbit sun
- Gravity provides attractive force
- Stable orbits (circular/elliptical)
- Empty space between bodies

Target: Atom

- Nucleus at center (massive, stationary)
- Electrons orbit nucleus
- Electromagnetic force provides attraction
- Stable orbits (energy levels)
- Empty space between particles

Mapping:

Sun ← corresponds to → Nucleus
Planets ← corresponds to → Electrons
Gravity ← corresponds to → Electromagnetic force
Orbital motion ← corresponds to → Electron orbitals

Transfer:

Solar system structure → Atomic structure
Orbital mechanics → Electron behavior
Stability conditions → Energy level quantization

Evaluation:

Success: Explained atomic structure
Limitation: Later quantum mechanics showed limitations
            (wave-particle duality, probability clouds)

But: Analogy was crucial first step, enabled progress

Analogical Retrieval Strategies

Problem: How to find relevant source cases?

Strategy 1: Surface Feature Indexing

Index by: objects, keywords, domain
Pros: Fast retrieval
Cons: Misses deep structural analogies

Strategy 2: Structural Indexing

Index by: relationships, causal patterns
Pros: Finds powerful cross-domain analogies
Cons: Computationally expensive, requires abstraction

Strategy 3: Hybrid

Initial filter: Surface features (fast)
Deep search: Structural similarity (powerful)
Balance: Speed and power

Strategy 4: Pragmatic Indexing

Index by: goals, constraints, functions
Retrieves: Cases relevant to current problem purpose
Example: All cases involving "transport liquid" goal

Design by Analogy

Engineering Design Example:

Target Problem: Design a new bicycle lock

Analogical Sources:

Source 1: Combination lock on safe

Mapping:
- Secure vault → Secure bicycle
- Combination → Code/pattern
- Multiple dials → Multiple rings

Transfer:
Design multi-dial combination lock for bicycle

Source 2: Key lock on door

Mapping:
- Secure house → Secure bicycle
- Physical key → U-lock with key
- Bolt mechanism → Lock shackle

Transfer:
Design U-lock with keyed cylinder

Source 3: Padlock with chain

Mapping:
- Secure storage → Secure bicycle
- Flexible chain → Cable
- Hardened lock → Reinforced lock body

Transfer:
Design cable lock with hardened padlock

Innovation:

Combine multiple analogies:
- U-lock shape (from door lock analogy)
- Combination mechanism (from safe analogy)
- Flexible cable extension (from padlock analogy)

Result: Hybrid design with multiple security features

Evaluation and Storage

After Transfer: Evaluate Solution

Success:

Store new case:
- Target problem
- Transferred solution
- Source analogy used
- Mapping details

Strengthen:
- Source-target connection
- Increase source retrieval probability for similar targets

Failure:

Analyze:
- Why did analogy fail?
- Was mapping incorrect?
- Was source inappropriate?
- Were there hidden differences?

Learn:
- Store as failure case
- Add constraints to prevent bad retrieval
- Identify limitations of source domain

Partial Success:

Iterate:
- Try different source
- Adjust mapping
- Combine multiple analogies
- Add domain-specific adaptation

3. Integration and Advanced Topics

EBL + Analogy

Complementary Methods:

Explanation-Based Learning:

Explains WHY solution works
Abstracts general principle
Enables transfer within domain

Analogical Reasoning:

Transfers solution structure
Enables cross-domain transfer
Maps relationships between domains

Combined:

1. Find analogous source (analogical retrieval)
2. Map source to target (analogical mapping)
3. Explain why source solution works (EBL)
4. Abstract explanation (EBL)
5. Transfer abstraction to target (combined)
6. Evaluate and store (both)

Example:

Target: How to reduce traffic congestion?

Analogy: Network packet routing (computer networks)
- Congestion in networks → Congestion in traffic
- Multiple paths → Multiple roads
- Dynamic routing → Dynamic traffic signals

EBL: Explain why dynamic routing works
- Detects congestion
- Redistributes load
- Balances utilization
- Adapts to conditions

Abstraction: "Congestion reduction requires:
- Real-time monitoring
- Load balancing
- Adaptive routing
- Multiple alternatives"

Transfer to Traffic:
- Install sensors (monitoring)
- Dynamic traffic signals (adaptive routing)
- Alternative route suggestions (load balancing)
- Real-time navigation apps (multiple alternatives)

Connection to Case-Based Reasoning

CBR vs. Analogical Reasoning:

Similarities:

Both retrieve past cases
Both adapt solutions
Both store new cases

Differences:

CBR:
- Within-domain transfer
- Surface + structural similarity
- Direct adaptation
- Many similar cases

Analogical Reasoning:
- Cross-domain transfer
- Structural similarity primary
- Deep re-interpretation
- Distant, dissimilar cases

Continuum:

Case-Based ←─────── Similarity ───────→ Analogy
(Same domain)                         (Different domains)
(Direct transfer)                     (Creative transfer)

Computational Challenges

Challenge 1: Retrieval

Problem: Find relevant analogies in large memory
Solution: Multi-level indexing (surface → structural)

Challenge 2: Mapping

Problem: Align source and target structures
Solution: Constraint satisfaction, structure mapping

Challenge 3: Adaptation

Problem: Transferred solution may not fit exactly
Solution: EBL-style abstraction, domain knowledge

Challenge 4: Evaluation

Problem: How to know if analogy is good?
Solution: Simulation, execution, expert judgment

Summary

Key Takeaways

Explanation-Based Learning extracts general principles from single examples by explaining WHY solutions work, abstracting essential features, and transferring to new situations. Requires rich domain knowledge but enables powerful generalization.
EBL levels: Specific example → Case-based (similar cases) → Explanation-based (abstract principle) → First principles (full generality). EBL operates at the practical middle level.
Analogical Reasoning transfers solutions across problems through four steps: Retrieval (find similar case), Mapping (align elements), Transfer (apply solution), Evaluation (test result).
Three types of similarity: Semantic (surface features, easy to recognize), Pragmatic (goal-relevance, functionally similar), Structural (relational patterns, most powerful for transfer).
Spectrum of analogy: Near transfer (same domain, direct application) to far transfer (cross-domain, creative insight). Far transfer requires structural similarity recognition.
Integration: EBL explains why analogies work, enables abstraction of transferred knowledge. Analogical reasoning provides the cross-domain bridge. Together they form powerful learning and reasoning system.

Essential Principles

Explanation enables transfer: Understanding WHY → Knowing WHEN to apply
Structure over surface: Deep relational similarity more valuable than superficial features
Abstraction finds essence: Remove non-essential details to reveal core principles
Cross-domain insight: Most creative solutions come from distant analogies
Knowledge required: Both EBL and analogy need rich domain models
Iterative refinement: Transfer, evaluate, adapt, repeat

Method Comparison

Method	Generalization	Transfer Distance	Knowledge Required	Examples Needed
Case-Based	Low	Near	Low	Many
EBL	High	Medium	High	One
Analogy	Medium	Far	Medium-High	One
First Principles	Highest	Any	Highest	Zero (deductive)

06 Advanced Reasoning