One of the key reasons I joined GA Tech and the Scheller College of Business is that I tend to draw on technology and business literature, and GA Tech is a great place for both. My current paper Exploration and Exploitation: An Essay on (Machine) Learning, Algorithms, and Information Provision draws on both these literatures. A key work on the idea of exploration versus exploitation in the business literature is James G. March, Exploration and Exploitation in Organizational Learning, 2 ORG. SCI. 71 (1989) which as far as I can tell has not been picked up in the legal literature. A good follow up to that paper is Anil K. Gupta, Ken Smith, and Christina Shalley, The Interplay Between Exploration and Exploitation, 49 ACAD. MGMT. J. 693 (2006). I had come upon the issue as a computer science question when working on a draft of my paper Constitutional Limits on Surveillance: Associational Freedom in the Age of Data Hoarding. That paper was part of my thoughts on artificial intelligence, algorithms, and the law. In the end, the material did not fit there, but it fits the new work. And as I have started to connect with folks in the machine learning group at GA Tech, I have been able to press on how this idea comes up in technology and computer science. The paper has benefitted from feedback from Danielle Citron, James Grimmelmann, and Peter Swire. I also offer many thanks to the Loyola University Chicago Law Journal. The paper started as a short piece (I think I wanted to stay at about five to eight thousand words), but as it evolved, the editors were most gracious in letting me use an asynchronous editing process to hit the final 18,000 or so total word count.
I think the work speaks to general issues of information provision and also applies to current issues regarding the way news and online competition work. As one specific matter, I take on the idea of serendipity which I think “is a seductive, overstated idea. Serendipity works because of relevancy.” I offer the idea of salient serendipity to clarify what type of serendipity matters. The abstract is below.
Legal and regulatory understandings of information provision miss the importance of the exploration-exploitation dynamic. This Essay argues that is a mistake and seeks to bring this perspective to the debate about information provision and competition. A general, ongoing problem for an individual or an organization is whether to stay with a familiar solution to a problem or try new options that may yield better results. Work in organizational learning describes this problem as the exploration-exploitation dilemma. Understanding and addressing that dilemma has become a key part of an algorithmic approach to computation, machine learning, as it is applied to information provision. In simplest terms, even if one achieves success with one path, failure to try new options means one will be stuck in a local equilibrium while others find paths that yield better results and displace one’s original success. This dynamic indicates that an information provider has to provide new options and information to users, because a provider must learn and adapt to users’ changing interests in both the type of information they desire and how they wish to interact with information.
Put differently, persistent concerns about the way in which news reaches users (the so-called “filter bubble” concern) and the way in which online shopping information is found (a competition concern) can be understood as market failures regarding information provision. The desire seems to be to ensure that new information reaches people, because that increases the potential for new ideas, new choices, and new action. Although these desired outcomes are good, current criticisms and related potential solutions misunderstand the nature of information users and especially information provision, and miss an important point. Both information users and providers sort and filter as a way to enable better learning, and learning is an ongoing process that requires continual changes to succeed. From an exploration- exploitation perspective, a user or an incumbent may remain isolated or offer the same information provision but neither will learn. In that case, whatever short-term success either enjoys is likely to face leapfrogging by those who experiment through exploration and exploitation.