SCEPTER: Strategic Chaos Engine for Planning, Tactics, Experimentation and Resiliancy

The Strategic Chaos Engine for Planning, Tactics, Experimentation and Resiliency (SCEPTER) program is developing analytic engines that can produce machine-generated strategies capable of competing with humans in the planning of real warfare, as evaluated within trusted simulation environments. 

SCEPTER technology can discover novel and surprising courses of action (COAs) by exploring the high complexity state-action space of military engagements at machine speeds. The high COA exploration speed is being enabled by tailorable abstraction of trusted, expert-informed models. A few of the best performing COAs will be validated in high fidelity trusted simulators and by human review. 

The program will explore two focus areas:

1. Development of unscripted goal-oriented agents capable of discovering novel, relevant, and interpretable COAs
   
   SCEPTER is developing methods capable of discovering unscripted and adaptive behaviors by using goal-oriented agents. These agents will operate in permissive, contested, and denied communication environments, and include the ability to discover and execute distributed kill-chains in such environments. Novel SCEPTER COAs will be validated against human experience as well as in trusted military simulators. Because humans are limited in the number of COAs that they can validate, this focus area is also developing methods and tools to assess the quality of the COAs against quantifiable measures. This ensures that the insights generated can be captured with a small number of human interpretable COAs.
    
2. Management of exponential growth of the global state-action space to achieve fast exploration of large-scale military scenarios
   
   This focus area will develop methods to manage exponential growth without unnecessarily limiting the exploration by the agents. To accomplish this, the SCEPTER program uses a wide variety of model abstraction methods, such as
   • providing a formal structure that defines work amongst agents
   • using linear combinations of small groups of agents
   • enforcing spatially local interactions in military scenarios
   • using causal or “narrative” interactions to identify smaller agent interaction groups
   • using methods to linearize the non-linear models of the agent interactions. 

Using these types of techniques, SCEPTER is able to achieve scenario exploration speeds of 10k-100k times faster than real-time in scenarios with up to 10,000 agents.