Statistical Debugging: A Hypothesis Testingbased Approach

1
Statistical DebuggingA Hypothesis Testing-based
Approach

• -Yiwei Zhang
• -Oct. 1st, 2007

2
Overview

• Addresses Fault Localization problem
• SOBER
• Dynamic
• no prior knowledge of the program semantics
• Just execution tracks with correct and
  incorrect labels
• Predicate-based
• Instrument the program with predicates (e.g.
  indexlt LENGTH)
• Record statistics information about evaluations
  of predicates for each executions.
• Hypothesis testing

3
A Motivating Example

A (m gt 0) B (lastm !
m)
Conclusion For a predicate, say A, the more
differently A is evaluated in correct and
incorrect executions of a faulty program, the
more likely A is fault-relevant
Faulty Program
Correct Program
4
Terminology

• Evaluation Bias
• Random Variable X
• Probability Model of
• Fault Relevant

5
SOBER
A collection of executions labeled with correct
and incorrect
6
Calculate the similarity of two probability models

• Hypothesis Testing
• null hypothesis
• Central Limit Theorem
• A random sample from
• A new random variable
• The new model of Y
• ,where

7
Calculate the similarity of two probability models

• Let be the probability density function.
• Given the observation of Y (calculated from a
  random sample of the incorrect executions), the
  likelihood of is
• A bigger likelihood of
• is more likely to hold
• a larger similarity between
• and

8
Evaluation

• Test Suite Siemens
• Methodology proposed by Renieris and Reiss
• A higher T-score means a lower localization
  accuracy.
• Compare with
• 7 existing fault localization algorithms.

9
Partial Result
10
The End