Posted by
A few months ago, I got a chance to talk with Thore Graepel (opens in new tab) of Microsoft Research Cambridge (opens in new tab) about Project Waterloo (opens in new tab), a Facebook game that is part of the Research Games (opens in new tab) project exploring behavioral game theory at a global scale.
GigaPath: Whole-Slide Foundation Model for Digital Pathology
Digital pathology helps decode tumor microenvironments for precision immunotherapy. In joint work with Providence and UW, we’re sharing Prov-GigaPath, the first whole-slide pathology foundation model, for advancing clinical research.
Now, Graepel and his teammates are at it again, this time with Doubloon Dash (opens in new tab), another game available on Facebook. While Project Waterloo enables players to allocate troops efficiently over a collection of battlefields, Doubloon Dash—a nautical scramble for a pirate’s buried treasure —offers a competition in which a pair of players have to invest, but only one can win.
“This game models situations like job interviews, patent races, or competitions,” explains Graepel, a principal researcher with the Machine Learning and Perception (opens in new tab) group at the U.K. lab. “In all of these cases, people are effectively playing an all-pay auction where everyone loses their bid but only one player gets the reward.
“The critical question then is: How much should you bet? If you bet more, your chances of winning are higher, but you are risking more money for a smaller reward. If you bet less, your chances of winning are smaller, but you risk less money and if you win, you win more. And then, there is your opponent. What will he or she do? Can you outsmart your opponent? The game, though, is less about beating your opponent and more about making the best of the situation—to earn as many doubloons as you can.”
The goal of the Research Games project is to test the behavior of real people in game-theoretic interactions, particularly those that occur in social networks.
“It is difficult to do such studies in a lab environment, as it is difficult to get a fair, bias-free sample of people,” says Pushmeet Kohli (opens in new tab), another researcher working on Research Games. “For instance, many past studies on capturing such behavior have used university students as participants, and their behavior might not reflect the behavior of the general population. Deploying these games on Facebook allows us to see the behavior of people from across the world.”
The nature of the new game presents players with an intriguing set of choices.
“With Doubloon Dash,” Graepel says, “we are hoping to learn how people behave in this strange type of all-pay auction. Are they careful and bid little, or are they reckless and bid a lot? How are they adapting their behavior to their opponents and over time?”
He also alludes to a “game-theoretically optimal solution” to Doubloon Dash, but researcher teammate Yoram Bachrach (opens in new tab) offers a slightly different take.
“I’d use the term ‘equilibrium strategy,’” Bachrach says. “We want to see not only if people deviate from the equilibrium behavior game theory predicts, but also how. If there are common deviations or used strategies, this means people are predictable and thus can be exploited.”
Such research, Graepel contends, could play a significant role within Microsoft.
“Studying such games in repeated settings,” he says, “can help us understand how trust and goodwill can overcome greed and mistrust to build fruitful long-term business relationships in which everyone is willing to make some concessions but both sides benefit in the long run.”