Assessing Video Game Level Design by Simulating Human Playtesters with Reinforcement Learning; The Technological Politics of Machine Learning in Criminal Justice

Razavi, Saeed, School of Engineering and Applied Science, University of Virginia
Vrugtman, Rosanne, EN-Comp Science Dept, University of Virginia
JACQUES, RICHARD, EN-Engineering and Society, University of Virginia

My technical project and my STS research paper find common ground in their discussion of machine learning. In my technical work, I propose a method for automatically testing how intuitively a video game level teaches its player how to play it, using reinforcement learning. In my STS research, I analyze the use and adverse effects of machine learning in the U.S. criminal justice system. I am dedicatedly interested in social justice, so beginning in STS 2500 Responsible Innovation during my third semester and continuing throughout my time doing computer science at UVA, I chose topics related to machine learning and the ethics surrounding it for projects in numerous classes. By doing this, I created an unofficial specialization in my learning and gained a greater understanding of the topics than I otherwise would have. With the application of machine learning as the common denominator, I chose my technical project to satisfy my lighthearted interest in game design as a hobby, while centering my STS research around a much heavier social justice related topic, which reflects my desire for a career in social justice and advocacy.
The technical portion of my thesis produced a method of automated video game playtesting that uses reinforcement learning agents to assess how effectively the design of a level in a video game can teach the player the mechanics of the game just through exploration of the level. Since game developers are familiar with the mechanics of their own games, it can be hard to playtest levels in the game as if the player were naive to the mechanics. As an alternative to hiring third-party playtesters, my method uses reinforcement learning agents as a stand-in for naive human players. My method relies on the analysis of how quickly a batch of agents learn to complete a level or meet a certain benchmark, compared to their performance on variations of the same level. The underlying logic in my design is that since humans and reinforcement learning agents explore and learn in generally similar ways, level configurations that are easier for agents to learn would also be more intuitive for human players. My method also involves imposing human-like limitations, like delayed reaction time, on agents so that their exploration and learning better represents that of a human without having to directly emulate real human behavior. My technical work will hopefully allow for more rapid prototyping to occur in small-scale game development teams without having to rely on third-party playtesters.
In my STS research, I analyze the arrangements of power that result from predictive policing algorithms used by law enforcement and risk assessment algorithms used in courtrooms. By analyzing how the design and impacts of predictive policing algorithms inherently encode and amplify marginalizing systemic biases, I determined that machine learning should not be used by police organizations in any way that attempts to predict criminality of civilians. I determined a similar conclusion in the case of risk assessment algorithms used in courtrooms by analyzing the way their designs learn and reproduce systemic bias, as the way that their lack of transparency harms legal rights to a fair trial. I determined, however, that if their decision logic is made more transparent with explainable artificial intelligence methods, risk assessment algorithms have the potential to contribute to more varied community outcomes. I ultimately determined that machine learning is unethical to use in criminal justice unless explicitly designed to counter structures of power which target and disempower marginalized populations.
Because I have been doing research into the domains of machine learning and machine learning ethics for more than half of my time at UVA, my technical project and my STS research acted as a culmination of my interest in the subjects. Through my technical project, I took the opportunity to get creative with machine learning, using some of what I have learned about the class of technology. In my STS research, I delved into the topics of social justice and particularly racial justice, in which I am personally very invested, by examining the ethical dimensions of machine learning in the context of criminal justice. By doing both of these projects together, I was able to engage with machine learning in both lighthearted and serious contexts, striking a balance for the sake of my own outlook on the technology. Through my work, I establish guidelines for the applications in which ML is and is not ethically appropriate to use, based on how different groups of people are affected. I hope going forward that ML engineers and government agencies alike consider the arguments that I put forth in my work, in addition to the large body of work advocating for more just use of ML technologies.
I would like to thank Professor Rosanne Vrugtman and Professor Richard Jacques for their invaluable guidance on my technical project and on my STS research respectively, as well as for the understanding, flexibility, and support that they gave me when I encountered personal setbacks while working on these projects. I would like to thank Professor Ben Laugelli for his support and guidance in writing my thesis project prospectus. I would also like to thank Professor Rider Foley for sparking my interest in technology-ethics in STS 2500 Responsible Innovation and for all of the encouragement and support he has shown me throughout my time at UVA. In addition, I would like to thank Professor Foley for being a genuine role model and one of my favorite professors with whom I have worked. Finally, I would like to thank and am incredibly appreciative of all of my friends and family who supported me throughout this work and my time at UVA. Whether by allowing me to talk through concepts with them, by peer reviewing and suggesting valuable revisions for my writing, or by providing emotional support to me when I dearly needed it, their love and support has been invaluable to me and I am more appreciative than I can possibly convey.

BS (Bachelor of Science)
Technological Politics, Predictive Policing, Risk Assessment, Machine Learning Ethics, Playtesting

School of Engineering and Applied Science
Bachelor of Science in Computer Science
Technical Advisor: Rosanne Vrugtman
STS Advisor: Richard Jacques

Issued Date: