Collaborating with machines
By using the sort of “cheap talk” that people use to strike up strategic collaboration, algorithms can learn to collaborate strategically with people.
In movies, robots often become less than friendly to their human creators. But in real life, advanced uses of machine intelligence, including driverless cars, autonomous drones, and autonomous stock trading will require machines to collaborate fully with people. Today’s machine intelligence is able to compete with humans and beat them at games like chess, poker, or Go, and it’s able to cooperate with people in simple situations where the algorithm and the person share a common interest. But to collaborate long term in real life, people work together strategically even when they have some competing interests, and machine intelligence has not yet been able to master this.
This paper from Nature Communications shows that this type of complex collaboration between machines and people is actually achievable, at least at the level of modeling called repeated, two-player stochastic games. The key is mastering a way people strike up collaborations that’s known in the field as “cheap talk.” In the future, machines could be taught to collaborate with people in even more complex and realistic scenarios.
Source: Nature Communications DOI: 10.1038/s41467-017-02597-8