The Defense Advanced Research Projects Agency (DARPA) Strategic Technology Office (STO) awarded Charles River Analytics a $9.9M contract to create automated strategies that can rival human expertise in planning complex operations. This Phase 2 work follows a $4M Charles River Phase 1 contract that demonstrated the success of a new neuro-symbolic AI approach for using strategic games to help teams plan complex operations.
Charles River, in collaboration with Data Machines Corporation (DMC), will further design and develop Meta-Reinforcement Learning System for Strategy and Tactics Assistance using Force-on-Force Forecasting (MERLINS-STAFF). MERLINS-STAFF is part of DARPA’s SCEPTER program (Strategic Chaos Engine for Planning, Tactics, Experimentation and Resiliency), which seeks to develop machine-generated strategies to help human planners.
Although AI has achieved remarkable success, such as roundly defeating top-ranked human players in an extensive variety of strategic games, it remains an outstanding research challenge to expand AI to support the complexities of real-world planning. Google DeepMind’s AlphaStar reached grandmaster-level performance in StarCraft II, but such AI solutions need to address specific shortcomings to meet the needs of more real-world planning problems, specifically dealing with longer duration plans, reducing training time, learning more complex and interdependent “moves,” and explaining what the AI knows.
DARPA’s SCEPTER program aims to push the boundaries of the state-of-the-art in AI planning. Current AI systems demonstrate expert-level creative play but only for short (approximately 15-minute) games with limited moves, and they require months to train using expensive, specialized hardware. In SCEPTER, DARPA is interested in a human-machine teaming system that can greatly expand the capacity of existing planning staff. The aim is not to replace human judgment but to enable planning staff to construct more high-quality, longer-duration plans, analyze plan strengths and weaknesses more deeply, and explore a larger number of plan alternatives in a fixed time.
Michael Harradon, Principal Scientist at Charles River and Principal Investigator on MERLINS‑STAFF, explains that real-time strategy (RTS) computer games serve as an ideal testbed for exploring complex scenarios, enabling progress measurement and testing of new AI techniques in a controlled environment. “In Phase 1, we achieved significant advancements, including playing longer games (e.g., greater than 1 day), reducing training time by 1000x, handling more complex entity controls, and creating human-understandable plans,” he says.
MERLINS-STAFF reached these milestones using innovative neuro-symbolic AI, combining machine learning with domain-specific knowledge for faster, more explainable AI models. This system collaborates with human planners, adapting to changing environments and producing creative, statistically significant improvements. “This provides an exciting foundation for our work in Phase 2 to more seamlessly team with human planners,” Harradon adds.
The new neuro-symbolic AI methods explored under this project are suited to several other real-world applications. For example, the underlying approach for MERLINS‑STAFF can be used for wide-ranging applications including climate simulation, physics-based modeling, or digital-twin modeling and simulation.
Contact us to learn more about MERLINS‑STAFF and our other capabilities in artificial intelligence and user interfaces & human-machine teaming.
