Original article can be found here (source): Artificial Intelligence on Medium
Toptal Scholarships for Women
I am a Data Scientist with a Masters degree in Econometrics and 4 years of experience in building and implementing robust machine learning models, both in academia and the private sector. My vision is to teach artificial intelligence (AI) how humans behave and how they make decisions under uncertainty; based on insights from behavioral economics, game theory and psychology and driven by AI’s extraordinary capabilities to model highly complex and dynamic relationships. I envision a world where policy makers can test their ideas by running large scale computer-simulations with agents that are based on actual humans and their real-world behavior. To achieve this goal, I will spend the next five years pursuing a PhD in Computer Science at the lab for Game Theory and Decision Theory, within the Artificial Intelligence Research Group at the University of British Columbia (UBC), Canada. With the Toptal Scholarships for Women, I will be able to pay the tuition for the first two years of my studies. In addition, I will profit from my mentor’s experience in how to communicate technological solutions effectively and how to encourage a broad audience, and policy makers in particular, to take advantage of them in their daily work.
Behavioral Economics x Artificial Intelligence
Predicting human behavior is ubiquitous in machine learning applications. Working as a Data Scientist, I realized that doing this successfully in real-world settings requires a deep understanding of the many biases that drive human behavior. Economists have traditionally modeled human behavior using the rational agent hypothesis, which states that individual agents make rational choices by maximizing expected utility. Behavioral economics has challenged this fundamental assumption by providing a rich body of experimental evidence on the many ways in which humans deviate from the rationality assumption. However, research into building general-purpose ‘behavioral’ artificial intelligence has been relatively scarce (see e.g. Wright, James R., and Kevin Leyton-Brown. “Level-0 meta-models for predicting human behavior in games.” Proceedings of the fifteenth ACM conference on Economics and computation. 2014.) and presents several challenges.
In particular, the exploration and validation of experimental evidence and resulting theories in the field of behavioral economics in real-world, empirical data has been challenging for two reasons; data scarcity and inadequate methodological paradigms. Data collection has only recently begun to take behavioral aspects into account. Where the data is available, traditional econometric approaches may not always be adequate to model more complex, non-linear patterns. I argue that ML/ AI methodology is ideally suited to tackle these challenges. Audio-visual and text data could be used to supplement existing data in order to address the scarcity issue. ML allows for complex, non-linear relationships and are thus well-suited to deal with these data types and make better use of existing numerical data. Moreover, pattern recognition and representation learning are core strengths of the field, and recent advances in unsupervised and few shot learning, causal machine learning and small sample methods open up other promising avenues. With my research, during my PhD and beyond, I want to tap into this area and help develop AI methods that address these challenges, as well as help to make the incorporation of behavioral aspects into applied settings feasible.
As outlined above, one of the most promising and far-reaching applications for behavioral AI is policy making. Uri Gneezy and Aldo Rustichini gave a pointed example of policy making gone wrong in their 2000 paper “A fine is a price” (Gneezy, Uri, and Aldo Rustichini. “A fine is a price.” The Journal of Legal Studies 29.1 (2000): 1–17.). In their study, the researchers examined what happened when a group of day-care centres introduced fines for parents who failed to pick up their children on time. According to the deterrence hypothesis, the introduction of a penalty should discourage the undesired behavior in rational agents, given everything else stays the same. However, Gneezy and Rustichini report that parents interpreted the penalty not as a fine, but as a fee for late pickup and as a result, the number of late-pickups increased significantly. Clearly, it would have been beneficial in this case to test the deterrence hypothesis with artificially intelligent systems that can simulate real-world human behavior and decision making. More broadly, general-purpose AI that can accurately simulate human behavior would be immensely beneficial for testing policies in quasi real-world settings in many domains of public policy making, including health, education or security.
To achieve this goal, I will spend the next five years learning the necessary skills during my PhD at the lab for Game Theory and Decision Theory at UBC. My background in machine learning and behavioral economics both from a theoretical and from an applied perspective makes me uniquely qualified to conduct research in this domain.
I am well suited to do this research because of my education as well as my past work experience in academia and the private sector. I studied behavioral economics and quantitative methods in my Masters, which helped build a profound theoretical understanding in both areas. I wrote my master thesis in this domain, using econometric methods to explore expectations and outcomes within the context of east-west migration in Germany. Specifically, I explored how over- and underconfidence differed across gender and industries, and how confidence and expectations affected migration decisions.
After my masters, I worked as a Data Science Consultant and experienced first hand the difficulties in building machine learning models that predict human decisions, e.g. in recommender systems aiming to predict purchasing decisions. Currently, I am working as a Data Scientist at the Centre for Gambling Research (CGR) at UBC. In this role I use machine learning to study behavioral patterns of problem gamblers on an online gambling platform and to build predictive models to identify problem gamblers. Results show that variability of money bet per session is the most important predictor of problem gamblers, which may be an indication of loss-chasing, a well studied behavioral pattern in the context of problem gambling. These results are summarized in a recently submitted paper (that I first-authored) to the gambling journal International Gambling Studies.
The Toptal Scholarships for Women will allow me to pay for the first two years of my education and will provide valuable insights when it comes to communicating my vision and how it can change the world to decision makers.