DRLFuzz

Installation

DRLFuzz requires the following dependencies:

• numpy
• pytorch
• PyGame-Learning-Environment
• pygame
• scipy

DRLFuzz experiments

catcher,flaappy_bird and pong are the three experimental subjects in the papper. They all contain three files:main.py,repair.py and verify.py.

• main.py: DRLFuzz script.
• repair.py: repair script.
• verify.py: verify the repair result script.

the result folder contains all experimental results.

• model: contain unrepaired and repaired models.
• xxx.log: experiment log.
• result_xxx.txt: generated failed test cases.
• result.xlsx: experimental analysis data.

STARLA experiments

STARLA algorithm includes two parts: generating test cases and execute them. The results are regarded as the executed ones. Main code files are as follows:

• STARLA.py: generate test cases.
• Execute_Results.py: execute results in Results folder.
• random_test.py: an example for using the environment.

Note: STARLA.py and Execute_Results.py are modified from STARLA under MIT license.

Additional Information.pdf

Additional Information.pdf introduces the replication experiments and results analysis of the STARLA method, applying STARLA to test trained agents on the Flappy Bird, Pong, and Catcher tasks. The experimental results showcase certain characteristics of STARLA and validate the limitations pointed out by the authors in their paper within the mentioned tasks.