Jiapeng Li "Reliable and Testable Deep Reinforcement Learning Systems"

HIIT Special Seminars
This talk is part of the HIIT Special Seminar series. The talks in this series are provided by candidates who have applied to our HIIT Fellowship recruitment call and are highly considered for the position. All talks are virtual, open to the public, and recorded for the future.
HIIT Special Seminar

When

–

Where

Online

Event language(s)

English
Join via zoom

This talk can be viewed via zoom. (Note: this talk will be recorded)

Title: 

Reliable and Testable Deep Reinforcement Learning Systems

Abstract: 

Deep Reinforcement Learning (DRL) systems have achieved remarkable success in domains such as robotics, game playing, and autonomous systems. However, the deployment of DRL in real-world applications remains limited due to concerns about reliability, robustness, and th lack of systematic testing methodologies. This presentation explores the challenges and emerging solutions for building reliable and testable DRL systems.

First, this presentation introduces mutation testing for DRL, which evaluates the effectiveness of test suites by injecting controlled faults into agents, environments, and training processes. It then presents metamorphic testing for DRL, addressing the oracle problem by defining metamorphic relations across inputs, environments, and policies. 

Building upon this, the presentation further introduces a tool framework for DRL metamorphic testing, which automates test generation, execution, and analysis. 

Finally, it discusses the emerging synergy between large language models and DRL, particularly in coding analysis tasks.

Overall, this work provides a comprehensive view of testing methodologies for building trustworthy DRL systems and outlines future directions for broader software systems.

  • Updated:
  • Published:
Share
URL copied!