Exploration versus Exploitation in Google's Think Quarterly

In Google's first issue of "Think Quarterly," it's new business to business publication, Susan Wojcicki, Google's employee number 16, sums up the classic exploration versus exploitation tradeoff writing, "We face the classic innovator’s dilemma: should we invest in brand new products, or should we improve existing ones?"

James March laid out this ubiquitous dilemma, which every organization faces in one form or another, in his now classic paper, "Exploration and Exploitation in Organizational Learning."  Each summer at the University of Michigan's ICPSR Summer Program on Quantitative Methods I co-teach a course on complex systems models in the social sciences in which I often discuss March's famous paper (in fact, we just discussed the paper today).  In going over the paper this summer I was struck again by the continuing relevance of his insights.

The quote that grabbed me today was, "... adaptive processes characteristically improve exploitation more rapidly than exploration ... these tendencies to o increase exploitation and reduce exploration make adaptive processes potentially self-destructive."  Here, March says we have to constantly be on guard to preserve exploration in our organizations.  Our natural tendency, just by doing what's best for us in the short run, is to gradually scale back exploration in favor exploitation, until all we do is exploit.  But, in doing so, we ultimately doom our organization to failure because we're no longer able to adapt to changing environment, or we lock into a sub optimal solution and eventually our competitors surpass us (see the earlier post on Borders).  March issued this warning to all organizations long before Clayton Christensen's Innovator's Dilemma.  The process of adaptation that makes us good at what we do now will destroy us down the road if we don't actively work to preserve exploration in our organization.  Which brings us back to Google.  Google is famous for so-called "20 percent time" in which engineers are asked to dedicate a full day a week to things "not necessarily in their job description."  This is Google's way of actively maintaining exploration in their organization.  So far, it seems to be working for them.

Diversity Trumps Accuracy in Large Groups

In a recent paper with Scott Page, forthcoming in Management Science, we show that when combining the forecasts of large numbers of individuals, it is more important to select forecasters that are different from one another than those that are individually accurate.  In fact, as the group size goes to infinity, only diversity (covariance) matters.  The idea is that in large groups, even if the individuals are not that accurate, if they are diverse then their errors will cancel each other out.  In small groups, this law of large numbers logic doesn’t hold, so it is more important that the forecasters are individually accurate.  We think this result is increasingly relevant as organizations turn to prediction markets and crowdsourced forecasts to inform their decisions.

Homophily and Information Spread

This article in Wired covers new research on networks and information by Sinan Aral (Northwestern B.A. in Political Science, MIT Sloan PhD, now at NYU Stern) and Marshall Van Alstyne.  The article describes research on the email communications of members of an executive recruiting firm, and says, “those who relied on a tight cluster of homophilic contacts received more novel information per unit of time.”  The article is confusing though because it mixes several distinct network concepts: homophily, strong ties, clustering, and “band width.”  Homophily is the tendency for people to be connected to other people that are similar to them; birds of a feather flock together. In his seminal paper, “The Strength of Weak Ties,” Mark Granovetter defined the strength of a tie as “a (probably linear) combination of the amount of time, the emotional intensity, the intimacy (mutual confiding), and the reciprocal services which characterize the tie”.  Clustering measures the tendency of our friends to be friends with each other.  And bandwidth is a less standard term in the social networks literature that captures the total amount of information that flows through a given tie per unit time (and thus is about the same thing as strength of a tie).

After reading the Wired piece, I’m left wondering if it is

  1. strong or “high bandwidth” ties through which we communicate a lot of total information,
  2. homophilic ties with people that are similar to us,
  3. ties with people that are members of a tightly knit cluster of friends, or
  4. all of the above

that provide us with the most novelty in our information diet.

A look at the original research article makes it more clear why the Wired article was so confusing.  The actual argument has a lot of moving pieces to it.  The first argument is that structurally diverse networks tend to have lower bandwidth ties.  Here structurally diverse appears to mean not highly clustered.  So, you talk more to the people in your personal clique than to people outside of your tightly knit group.  The second piece relates structural diversity to information diversity.  They find that the more structurally diverse the network, the more diverse the information that flows through it.  So far, this seems to line up with the standard Granovetter weak ties story.  The third relationship is that increasing bandwidth also increases information diversity, and more importantly, increasing bandwidth increases the total volume of new (non-redundant) information that an individual receives.  The idea here is that if you get tons of information from someone, some of it is going to be new.

Finally, since both structural diversity and bandwidth increase information diversity, but structural diversity decreases with increased bandwidth, they set up a head to head battle to see whether the information diversity benefits of increasing bandwidth outweigh the costs of reducing structural diversity.  They have three main findings on this front that characterize when bandwidth is beneficial:

  • “All else equal, we expect that the greater the information overlap among alters, the less valuable structural diversity will be in providing access to novel information.”
  • “All else equal, the broader the topic space, the more valuable channel bandwidth will be in providing access to novel information.”
  • “All else equal, ... the higher the refresh rate, the more valuable channel bandwidth will be in providing access to novel information.”