DARPA Wants Wargame AI To Never Fight Fair

ALBUQUERQUE: Northrop Grumman is building an AI designed to find new strategies to break virtual opponents. Future AI tools, based on this research, could help human commanders break opponents in real battles.

The contract is part of DARPA’s Gamebreaker program, which wants to turn the design considerations of modern strategy games on their head, using AI to find every unfair advantage hidden in the game.

“Gamebreaker seeks a methodology for finding “broken states” in games – situations in which one player in the game can gain unexpected advantages over a competitor,” Joshua Bernstein, director of advanced intelligent systems at Northrop Grumman, says. “In these applications AI finds asymmetrical conditions in a system (eg, the game or a real-world scenario) and communicates these conditions to stakeholders, such as military planners.”

Gamebreaker program is focused on a range of real-time strategy games, or programs where players command a range of units with different characteristics in competition against each other. These include the popular Starcraft series of games, where players build starfighters, train marines, and extract minerals in a futuristic alien setting. It also includes games like Google Research Football, which uses a physics engine to model physical impacts of virtual competitors players in a soccer game.

Northrop Grumman’s entry will be built in Command: Modern Operations, a hyper-realistic theater-wide combat simulator designed to model Cold War as well as present conflicts.

“If we can figure out a generic method to assess and then manipulate balance in commercial video games, my hope is that we might then apply those AI algorithms to create imbalance in DoD simulated war games used to train warfighters for real-world battle,” Lt. Col. Dan “Animal” Javorsek, the Gamebreaker program manager in DARPA’s Strategic Technology Office, said when the program was announced in May 2020.

Gamebreaker is, especially in terms of Pentagon budgets, an almost minuscule contract, clocking in at just $1 million. Its focus on developing an AI that can win scenarios in one game, and then testing if that AI can win a second game, is somewhat narrow. Yet the implications for more accurate wargaming through thoughtful AI could have a huge impact on how weapons systems are designed, modeled, and ultimately used by human commanders aided by AI agents.

“Wargaming is a well-established and critical element of real-world military planning and weapon system development, especially in complex scenarios such as those our users intend to address with concepts such as JADC2,” said Bernstein. “Gamebreaker is focused on the development of a methodology to explore and identify unique opportunities to defeat an adversary in competition – while we are developing those tools using games, we believe Gamebreaker’s methods will have direct applicability to the military services’ development and employment of joint all domain operational concepts.”

Gamebreaker, part of DARPA’s larger initiative in military AI, is about winning Real Time Strategy games. Combining entertainment with simulation, these games seek to foster both fun and a balanced, competitive experience, one where each player stands a reasonable chance of winning.

This is directly at odds with actual war, where the smoothest path to victory is maximizing every unfair advantage a side can muster against a rival.

To succeed at breaking this balance, teams must build an AI that can play a strategy game, and then, while staying within the rules of the game, figure out how to use all the available pieces in the best and most unfair way against its opponents. It is about novel tactics, without any of the limitations of human understanding holding back how the algorithm plots a path to victory.

“Command offers considerable flexibility for manipulating the capabilities of the simulated combat units, and importantly, the game builds these units based on real-world systems,” said Bernstein. “As a consequence, Command allows us to explore the implicit beliefs a weapon system developer had when designing a given system, allowing us to experiment with novel approaches to various real-world mission scenarios.”

Command: Modern Operations is built on open-source data. That means Northrop Grumman can easily replace the game’s default information with more accurate characteristics, like undisclosed missile speeds or sensor ranges.

For some scenarios, like a Cold War game where Warsaw Pact naval forces use submarines to launch an attack against NATO patrols in the Arctic sea, some of the information about vehicle and weapon capabilities is public and declassified and already incorporated into the game. In more modern scenarios, such as a hypothetical battle in the South China Sea, the game incorporates an open-source understanding of the technical capabilities involved, and the open-source code allows users to update it with more publicly unavailable information.

The open-source code also allows Northrop Grumman to easily insert their own AI agent into the game, and study how it interacts with all the available units. These include systems as large as aircraft carriers down to the precision of guided missiles launched from helicopters. Crucial to how Command: Modern Operations works is the way it models and incorporates sensor data, revealing submarines only where they were last observed, rather than showing them on-screen if they are out of the reach of any sonar systems.

If there exists a novel strategy in the game that allows a commander to more effectively employ, say, ship-launched guided missiles and helicopter-borne sensors, the goal is that the Gamebreaker AI will find it. For initial testing, the Gamebreaker AI will only compete against other AI players, though the potential exists for it to play against human opponents.

Proving that AI can win games is an essential first step to proving that AI can actually offer useful insight to commanders. If humans are going to place trust in an algorithmic order of battle, it is absolutely essential to have faith that the algorithm knows what it is doing.