Thirty-five years ago, while Martin Amis was writing “Money,” one of the novels that defined the nineteen-eighties, he admitted to a distracting dalliance with another contemporary icon. “I have spent weeks in a PacMan-fed stupor, unwilling and unable to think about anything else,” he wrote in “Invasion of the Space Invaders,” his “addict’s guide” to the nascent arcade. Amis was not alone in his obsession. The Japanese-made game, in which players guide an auto-munching yellow head through a Daedalian maze, consuming a trail of pellets while fleeing four candy-tone ghosts, earned more than a billion dollars in quarters in its first year, surpassing the highest-grossing “Star Wars” film at the time. Pac-Man towered, Amis wrote, over “a vast garbage dump of rocky romances and wrecked careers.” He claimed to know a young actress with a case of Pac-Man hand so severe that her index finger resembled a piece of liver.
But Pac-Mania was short-lived. The way in which the game was programmed meant that skilled and attentive players could eventually memorize its patterns. With mastery came boredom. So, in 1981, a group of M.I.T. dropouts decided to modify the code with new maze layouts and an adapted artificial-intelligence routine, designed to make the game more surprising and challenging, and therefore more profitable. The plan was to sell Crazy Otto, as it was initially known, as a kit to arcade owners, which they could use to modify their forsaken Pac-Man cabinets. But the team’s reworking of the game proved so compelling that Midway, Pac-Man’s U.S. distributor, bought the design. The character’s yellow head was given a red bow, scarlet lipstick, and a glamorous beauty spot, and, lo, the Man became a Ms.
Ms. Pac-Man’s grand innovation was to randomize the movements of the ghosts, named Inky, Blinky, Pinky, and Sue. (The last of these used to be called Clyde; we were never told what befell him.) In the original game, when Pac-Man eats a piece of points-giving fruit, the ghosts turn a horrified blue and, in this temporarily vulnerable state, flee to one of the maze’s four corners. “Rather than sending each ghost to their favorite corner, our algorithm randomly picks a corner for each monster each time,” Steve Golson, one of Ms. Pac-Man’s developers, told me. In addition, the fruit itself, which was static in the first game—“there simply to tempt you into hubristic sorties,” quoth Amis—was given the gift of movement. Now every levitating strawberry, pear, banana, and cherry would wend its way through the maze unpredictably, along one of sixteen possible paths. Players could no longer test and perfect the optimal route; now they had to navigate on the fly.
Three decades later, these simple tweaks serve a weighty new purpose. The skill that Ms. Pac-Man demands of its players—making multi-objective, dynamic decisions quickly—turns out to be the same ability that artificial-intelligence researchers wish to program into many of their bots. And according to Silvia Ferrari, the director of Cornell University’s Laboratory for Intelligent Systems and Controls, the game is an especially ideal environment for training autonomous military machines. In January, a bot that she created with three of her colleagues set a new A.I. world record for Ms. Pac-Man, beating the previous record by more than seven thousand points. The program defines its strategy based on the shape of the maze, the locations of the pellets, and mathematical models of the ghosts’ behavior. While the endeavor may seem flippant, Ferrari, who first played Pac-Man when she was ten years old, believes that it will lead to useful applications. Code tested and trained using Ms. Pac-Man could, she suggested, be integrated into unmanned vehicles, helping them conduct search and surveillance missions under conditions that would be too hazardous for humans—in war zones, disaster areas, or the deep ocean. “Ms. Pac-Man makes for an excellent benchmark for robots, autonomous systems, drones, and mobile sensors,” she said.
A.I. researchers have been interested in games for years. As Mike Traweek, Ferrari’s program officer at the Office of Naval Research, told me, “Games are affordable and, with their formally described rules for scoring, winning, or losing, are often a powerful and convenient way to provide an objective context within which to balance theory and practice.” More than that, though, they are a cheap means of giving bots the huge repository of data—moves, countermoves, strategies, and common outcomes—that they require to learn. “Data are expensive,” Traweek said. “Statistically meaningful data are often prohibitively expensive.” Oriol Vinyals, a scientist at Google DeepMind, told me that the beauty of games is that “they offer limitless training data and measurable progress, and can also be customized to emphasize different tasks, such as planning or navigation.” Vinyals worked on AlphaGo, which last March became the first computer program to beat a professionally ranked human player of Go. Just as Ferrari’s Ms. Pac-Man program could yield advances on the battlefield, Vinyals claims that techniques discovered through the creation of AlphaGo have already led to a forty-per-cent reduction in the amount of energy used for cooling Google’s data centers.
Since 2000, when Congress mandated that a third of military ground vehicles and so-called deep-strike aircraft be replaced by robotic vehicles within a decade, the defense industry’s interest in A.I. has steadily grown. One possible implementation of Ferrari’s program, Traweek said, could be in a field that closely resembles the hazard-laced maze found in the game: mine disposal. “In Ms. Pac-Man, there are opportunities to use prior experience and prior knowledge in a way that delivers more expectation for success than randomly wandering around,” Traweek told me. Where a typical program might search an area methodically, section by section, until the probability of finding a mine had dropped to an “acceptable level,” he said, Ferrari’s would use “every scrap of reliable prior information to allow me to design a search strategy that allows me to reach the same degree of confidence sooner.”
For the moment, though, this is more speculation than prediction. The Navy has been using autonomous systems to control submarine-launched cruise missiles for years. But something as complex as an A.I. minesweeper, Traweek said, “is a problem so far removed from balancing a rocket on a nozzle that it demands a fundamentally new definition of control.” Arcade games are one of the ways in which that definition is being solidified. Still, as the threat of litigation from the families of those wounded and killed by autonomous machines grows—earlier this month, the husband of a factory worker killed in Michigan by a wandering robot sued five companies involved in the device’s creation—it seems possible that Ms. Pac-Man could one day be called to perform a role that her creators never imagined: providing evidence in a court of law.
