Automated Program Repair Tool (GenProg-Style)

Java implementation of a GenProg-style Automated Program Repair (APR) engine for Task 03.

Team

• Group 7
• Tim Greller
• Jonas Heitz
• Rafail Venediktov

1. Overview

This tool takes a buggy Java program, a test suite, and fault-localization weights, then searches for a patch using a genetic programming loop.

High-level flow:

1. Load benchmark metadata and source files.
2. Build initial patch population.
3. Evaluate candidates by compile + test execution.
4. Select better candidates.
5. Generate offspring with crossover + mutation.
6. Repeat until all tests pass or limits are reached.

2. GenProg Mapping

Implemented pipeline is assignment-oriented and close to GenProg pseudocode:

• weighted fault localization guides mutation target choice
• patch population is evolved over generations
• fitness is test-based
• stop criteria: full test pass, generation cap, time cap

Core GenProg edit operators:

• DELETE
• INSERT
• SWAP

Additional generic expression-level operators (for practical Java repairability):

• REPLACE_EXPR
• MUTATE_BINARY_OPERATOR
• NEGATE_EXPRESSION

3. Fault Localization Policy (Strict)

The current repository behavior is strict with respect to FL weights:

• suspiciousness values come only from fault-localization.json
• expected levels are 1.0, 0.1, 0.0
• no fallback exploration of 0.0 statements in statement selection
• no suspiciousness normalization in selection probability

4. Project Structure

.
├── src/main/java/edu/passau/apr/
│   ├── Main.java                      # CLI entrypoint, run orchestration
│   ├── algorithm/
│   │   └── GeneticAlgorithm.java      # Evolution loop (init/select/crossover/mutate/evaluate)
│   ├── config/
│   │   └── Config.java                # Runtime parameters + defaults
│   ├── evaluator/
│   │   └── FitnessEvaluator.java      # Compile + JUnit execution + fitness calculation
│   ├── model/
│   │   ├── BenchmarkConfig.java       # Parsed benchmark metadata
│   │   ├── Edit.java                  # Edit representation
│   │   ├── FitnessResult.java         # Fitness tuple/state
│   │   ├── Patch.java                 # Patch object + mutation/apply logic
│   │   └── StatementWeight.java       # FL weight model
│   ├── operator/
│   │   ├── PatchGenerator.java        # Random/guided patch generation + crossover
│   │   └── PatchGeneratorTest.java    # Generator-focused tests
│   └── util/
│       ├── AstUtils.java              # Shared AST safety/compatibility checks
│       ├── BenchmarkLoader.java       # JSON loading for benchmark/fl config
│       └── Pair.java                  # Utility pair
├── benchmarks/                        # Benchmark suite B01..B12
├── docs/
│   ├── assignment/                    # Task PDF + declaration template
│   └── papers/                        # GenProg papers
├── test_quick.sh                      # Sweep script for all benchmarks
├── README.md
└── REPORT.md

5. Benchmarks

Current benchmark set:

• B01_OffByOne
• B02_DuplicateLine
• B03_MissingStatement
• B04_WrongPredicates
• B05_WrongIndex
• B06_InventoryNormalization
• B07_IntervalPlanner
• B08_LogReportBuilder
• B09_ShippingTierSwitch
• B10_BillingProration
• B11_ConsecutiveBatcher
• B12_TagSanitizer

Each benchmark directory contains:

• buggy/ — buggy source
• fixed/ — reference fixed source
• tests/ — JUnit tests
• benchmark.json — paths + test class names
• fault-localization.json — suspiciousness weights

6. Requirements

• Java 17+
• macOS/Linux shell (examples use bash)
• Gradle Wrapper included (./gradlew)

7. Build and Run

Build

./gradlew build

Run one benchmark

./gradlew run --args="--benchmark benchmarks/B01_OffByOne --seed 42 --maxGenerations 80 --timeLimitSec 120" --no-daemon

Run full sweep

./test_quick.sh
# or explicitly:
./test_quick.sh 42 80 120

8. CLI Options

• --benchmark <path> required
• --seed <long>
• --maxGenerations <int>
• --timeLimitSec <long>
• --populationSize <int>
• --positiveTestWeight <double>
• --negativeTestWeight <double>
• --mutationWeight <double>
• --runTests <buggy|fixed> (compile+run tests only)
• --verbose

Defaults are defined in /Users/rafailvv/Учеба/University of Passau/Program Repair/Task 3/src/main/java/edu/passau/apr/config/Config.java.

Run Tests Only (Buggy vs Fixed)

./gradlew run --args="--benchmark benchmarks/B01_OffByOne --runTests buggy" --no-daemon
./gradlew run --args="--benchmark benchmarks/B01_OffByOne --runTests fixed" --no-daemon

9. Output Layout

Patched source output:

• out/<benchmark>/patch_<timestamp>/<MainClass>.java

10. How to Read test_quick.sh Output

• Status: OK means the APR process exited with code 0.
• Actual repair success is indicated by:
  • SUCCESS: Found a patch that passes all tests!
• No solution found within limits. means no full repair was found under given search budget.

11. Common Troubleshooting

• If runs are unstable between launches, keep seed fixed and increase --maxGenerations / --timeLimitSec.
• If runs look stuck, verify FL file points to correct buggy lines.
• If output is very verbose, redirect stderr or use filtered script output.
• If Gradle daemon state is noisy, use --no-daemon as in examples.

12. Notes

• APR is stochastic by design; outcomes can vary by seed and budget.
• The implementation remains benchmark-agnostic at algorithm level (no benchmark-specific hardcoded fixes).