Applying Computational Game Theory AI to Many-v-Many BVR Air Combat

Computational game theory is the only sound way of planning against an intelligent adversary: it is the only approach that can properly handle imperfect information (partial observability) and the associated opportunities to deceive and be deceived, as well as a nonstationary intelligent adversary that selects its strategy based on what strategy we select for ourselves. Strategy Robot (SR) has developed the only game-theoretic COA Generation and Execution System (CGES). It computes COAs for Blue and Red simultaneously, taking into account that Blue’s optimal COA depends on Red’s COA and that Red’s optimal COA depends on Blue’s COA. So, unlike simulation, single-sided optimization, and single-sided reinforcement learning, it does not need to assume that the user has correctly guessed the adversary's COA (but it can further capitalize on aspects of it being known). Our computational game theory AI has already generated superhuman COAs for armored combat and no-limit Texas hold’em poker. We propose to apply CGES to BVR air combat. The approach in CGES is based on game-theoretically constructing COAs from COA pieces (COAPs). This enables the planning abstraction to be raised so strong COAs can be generated with a reasonable amount of computation. The proposed work includes generating COAPs for air combat for a series of increasingly complex scenarios. Strategy Robot’s world-leading AI experts will work closely daily with Strategy Robot fighter pilots (some of the best in the nation) to iteratively improve the COAPs, COA constraints, utility functions, and so on. The work will include computational AI training and thorough incremental evaluation of COAs in simulation and live flight. The work will go all the way to the TS/SAP level and Strategy Robot can fully support that. The programming will be done in Pittsburgh, with close in-person regular interactions with Eglin AFB.