Projects

Page under construction

Please see https://github.com/nphamilton to see some of the projects I have worked on in the last couple of years. I will be adding detailed descriptions soon.

Research Summary

During my time with the Verification and Validation of Intelligent and Trustworthy Autonomy Laboratory (VeriVITAL), my research has focused on the development of safe and robust Machine Learning (ML) controllers for various Cyber-Physical Systems (CPS). In my studies, I found Reinforcement Learning (RL) to be the most promising ML approach. Thus, most of the work listed below has to do with RL.

stl-gym

A tool for monitoring Signal Temporal Logic (STL) specifications in OpenAI Gym environments and replacing the reward function with the degree of robustness.

This project is still under development.

mlv_2020_project

A project focused on uncovering how safe Safe Reinforcement Learning approaches are by formally verifying the trained network.

This project inspired many of my later works in comparing different Safe Reinforcement Learning approaches. Here, I wanted to formally verify the trained networks created using a method from an award-winning paper. In the process of re-creating their results, using their code, I found inconsistencies in how they compared their approach to previous ones. When I accounted for these inconsistencies, I found all the claims that their approach speeds up the training process were unfounded. Instead, all the performance gains were the result of simplifying the tasks for their agents.

I outlined everything I did in this write-up and will link the blog post describing a lot of the takeaways when I’ve finished writing it.

rl_library

A collection of RL algorithms and training applications I wrote.

I started this project in an attempt to create understandable RL algorithms. I started this before stable-baselines made a commitment to maintaining good documentation and before SpinningUp was released. The main advantage of continuing to use these implementations is when training agents in ROS applications.