Most military settings are imperfect-information games. They are games not in the sense of being recreational but in the sense that there is at least one adversary. They are games of imperfect information in that there is fog of war: the player does not know the state of the world exactly—for example, an adversary’s resources, capabilities, locations, and readiness. Imperfect-information games are much more difficult than perfect-information games, and algorithms for perfect-information games (such as Deep Blue, AlphaGo, AlphaZero, and MuZero) do not apply at all because they cannot address issues such as deceiving effectively, understanding deception by others, and reveal/conceal. [1]

Revolutionizing Military Planning with Advanced Strategies

Today, military planning settings are not treated in a sophisticated way as imperfect-information games. Instead, in military planning, war gaming, command and control, and doctrine generation, plans (that is, strategies in game theory terminology) are generated manually based on gut feel, with simulation or table-top exercises sometimes as a support tool. In military settings, Red and Blue typically have more possible strategies than the number of atoms in the universe. Yet only a tiny number of strategies for Red (typically 1-5) and for Blue (typically 1-100) are tested. So, only a vanishingly small portion of the strategy space is evaluated, causing enormous risks and overestimation of the strength of Blue’s strategies (because Red’s strategy space is not fully explored) and leaving significant opportunities on the table (because Blue’s strategy space is not fully explored). Also, today’s military plans are typically deterministic, while randomization is sometimes needed to make strategies non-exploitable (as a simple example, consider Rock-Paper-Scissors). Furthermore, hiring people to play Red in training, and hiring humans to do planning, is costly and slow.

In much of DoD’s planning today, attention is paid to Blue’s logistical constraints, timing, etc., but very little attention is given to Red’s strategy, which affects Blue’s outcomes. The weaknesses of that approach have only been exposed to a limited extent because we have been fighting against adversaries that are much smaller and have drastically weaker assets. However, now with attention appropriately shifting to peer adversaries, that approach is woefully inadequate.

Pioneering Superhuman Decision Making in High-Stakes Settings

In contrast, Strategy Robot’s computational game theory tools compute optimal—and usually novel—strategies for Blue and Red simultaneously, taking into account that Blue’s optimal strategy depends on Red’s strategy and Red’s optimal strategy depends on Blue’s strategy. This is called strategic reasoning in game theory because one side has to think about the other side’s strategy in order to intelligently decide what to do. So, unlike in simulation, the players’ strategies are output rather than input. There is significant opportunity to improve planning, replanning, training, and command and control using modern computational game theory AI techniques. This kind of AI technology can also be used to evaluate given strategies for Blue and/or for Red, regardless of whether the strategies were generated by hand or computationally. These techniques can thus also be used to evaluate and train officers playing against them.