Outline
Logistics
Defining a Learning Problem
Concept Learning
Evaluating Attributes
Resulting Tree ….
Summary: Learning = Search
Correspondence A hypothesis = set of instances
Version Space: Compact Representation
Training Example 3
Comparison
Two kinds of bias
PAC Learning
Ensembles of Classifiers
Constructing Ensembles
Review: Learning
Softbot Perception Problem
Strategy: Wrappers
Scaling issues
Wrapper Induction
Example
LR wrappers: The basic idea
Country/Code LR wrapper
Observation
Ubiquity!
Inductive (example-driven) learning
Wrapper induction algorithm
Step 3: Finding an LR wrapper
LR: Finding r1
LR: Finding l1, l2 and r2
Finding an LR wrapper: Algorithm
A problem with LR wrappers
The complication
A solution: HLRT wrappers
Country/Code HLRT wrapper
“Generic” HLRT wrapper
Step 3: Finding an HLRT wrapper
HLRT: Finding r1, l2 and r2
HLRT: Finding h, t, and l1
Finding an HLRT wrapper: Algorithm
Step 1. Termination condition
PAC model
PAC model for HLRT
PAC model: Interpretation
Step 2. WIEN: Manual page labeling
Automatic page labeling
Recognizers
Corroboration of Imperfect Recognizers
Corroboration: Example
Summary of results
Q: Is wrapper induction practical?
A: Yes
Kushmerick Contributions
MDP Model of Agency
Trajectory
MDP Model (continued)
Good News and Bad News
Properties of the Model
Computing Optimal Policies
Policy Construction and Dynamic Programming
Value Iteration and Its Variants
Policy Iteration
Summary of MDP Solution Techniques
Reinforcement Learning
Q Learning
Q Learning (cont.)
Convergence of Q update
Summary of General MDP Model
Summary of Reinforcement Learning
PPT Slide
Simple Backup
Email: weld@cs.washington.edu
Other information: CSE 592, Lecture 9
Download presentation source