OUSD (R&E) CRITICAL TECHNOLOGY AREA(S): Trusted AI and Autonomy; Advanced Computing and Software OBJECTIVE: Implement a game server capable of engaging significant portions of allied warfighters in operationally relevant, enjoyable, and analyzable joint operation wargames. This wargaming cloud will harness the American democratic and competitive ethos to both train our service members in the operational warfighting family business and crowdsource the development of potentially disruptive operational strategies. The dataset created through this effort will enable both traditional data analysis methods and more modern approaches based on machine learning and artificial intelligence. SBIR phases will seek a warfighting game that balances playability, DoD relevance, and data extraction capability. DESCRIPTION: Our proposal aims to build an initial repository of operational wargames designed to educate every allied warfighter on the intricacies of the operational level of war while enabling statistical analysis and AI-informed decision-making through significant quantities of game iterations. Leveraging public and popular games, the SBIR awardees will produce a large dataset from their existing online servers from which military planners could derive decision analyses at the appropriate operational level. Further, if not already developed, SBIR funds should develop a Markov Decision Process dataset for reinforcement learning applications. Additionally, SBIR awardees will provide plug-in capabilities to their games which allow for the DAF to adjust and create novel scenarios and assets. These plug-ins will provide extensibility and adaptability for future in-depth data-driven strategy analysis. Further, utilizing these gaming platforms, our approach will allow every airman and guardian to test their operational instincts against the best tacticians worldwide, fostering a sense of pride, competition, and ownership while teaching the family business of warfare. The datasets that are created via these games will populate a gameplay database, which can be used to analyze trends from worldwide player data, develop alternative strategies from that data, and train AI agents. These trained AI models will enhance the high-level traditional wargaming process in three primary ways. First, it will add fidelity to adjudication by actually simulating tactical level encounters based on moves, rather than the current process of having white cell' declare an outcome based on a spreadsheet, dice-roll, or rule of thumb. Second, it will greatly accelerate logistics and laydown planning, which provides re-playability. One initial early finding from the adoption of Command:PE was that human planners only started taking risk and exercising creativity after the conventional' plan had tried and failed multiple times, but when they did they were able to actually start winning scenarios that were assumed losses. Replayability gets human players into a place where they can produce these valuable outcomes - if the work-hours required to run a traditional wargame only allow for one rep, bold concepts and disruptive approaches may not get a hearing. Last, AI agents can provide the ability to MoneyBall' diverse approaches to wargaming and planning. Anti-fragile' strategies that incorporate both chaos and order is a strong suit for a free society, especially against an authoritarian regime. Since logistics planning is a necessity, this form of modeling would allow for enough branches to make space for mission command at echelon, which will in turn impose costs on an adversary well prior to conflict. In order to pursue these objectives, three goals must be met by the AI modeling effort: -1) Tactical modeling. The ability for AI agents to model tactical encounters in a relevant wargaming system, which will provide a rigorous tool for adjudicating operational-level wargaming moves. -2) Logistical modeling. Given a combat desired force in a scenario, AI agents can model one or ideally several scenarios for basing and logistical support. -3) Operational modeling. (stretch goal) Given an operational design, complete laydown, tactical encounters, and operational level branching in order to provide a strawman' initial analysis of a concept of operations. This proposal supports the requirement for DAF warfighters to be educated about real operational threats. Further, this will provide the ability for warfighters to better assess strategies, tactics, and procedures against thinking and adaptive opponents. In so doing, this SBIR will help prepare DAF members to be ready to deploy and fight (OI 7) while ensuring an operational understanding of JADC2 (OI 2) and enabling the analysis of alternatives for resilient forward-basing options (OI 5). Engaged Stakeholders: AFWERX Spark, AFIMSC, Morpheus, Air Force Gaming, Lincoln Labs, DAF, MIT AIA PHASE I: The objective of Phase I is that projects will demonstrate their game's playability, DoD relevance, and data extraction capability. The team is seeking games that can balance abstraction and realism, sufficiently mimicking the operational level of war for warfighter education and human evaluation while maintaining high levels of engagement and playability. Additionally, games will demonstrate their ability to export gameplay data that fully and efficiently captures in-game experience for a broad gamut of post processing. In this feasibility study, companies will demonstrate their capability of data extraction. PHASE II: Phase II will focus on game flexibility, scaleability, and capability demonstration with real gameplay. In addition to Phase I goals (playability, relevance, and extraction capability), performers will demonstrate the commercialization potential of their game (more data to capture for AI agent training), their ability to host their game on government servers and provide a continuous stream of data during the PoP from all hosted games. Additionally, performers shall give the USG the ability to extend scenarios with user-defined assets, inject AI agents as players, and permit faster-than-real-time command-line gameplay suitable for agent training. PHASE III DUAL USE APPLICATIONS: The future of gaming will require extensible, AI-ready games capable of employing cooperative and competitive agents as NPCs. The ability to inform the development of these agents using real gameplay data from experienced users could be invaluable. These capabilities for small game companies improve the reach, enjoyability, and accessibility of their games to the worldwide market. In ensuring their games are AI-ready, games will improve their marketability for future research and development to unique markets. REFERENCES: Vinyals, Oriol, et al. "Grandmaster level in StarCraft II using multi-agent reinforcement learning."Nature 575.7782 (2019): 350-354.; Meta Fundamental AI Research Diplomacy Team (FAIR) , et al. "Human-level play in the game of Diplomacy by combining language models with strategic reasoning." Science 378.6624 (2022): 1067-1074.; Siu, Ho Chit, et al. "Evaluation of human-AI teams for learned and rule-based agents in Hanabi. "Advances in Neural Information Processing Systems 34 (2021): 16183-16195.; Lyons, Joseph, et al. "Measuring Perceived Agent Appropriateness in a Live-Flight HumanAutonomy Teaming Scenario." Ergonomics in design (2022): 10648046221129393.; Silver, David, et al. "Mastering the game of Go with deep neural networks and tree search." nature 529.7587 (2016): 484-489.; KEYWORDS: Wargaming, Data Analysis, Artificial Intelligence, Imitation Learning, Reinforcement Learning
