Forecasting and Decision Impact Analysis from Ripple Effects of Behaviors

B. Frutchey,
United States

Keywords: forecasting, Decision Intelligence, Tensor Completion, Machine Learning, Driver Analysis, Scenario Simulation


Every action sets in motion ripples of effects that spread throughout an environment. This simple fact empowers analysis, especially in situations where the action of interest cannot be directly observed. For our military, adversaries trying to hid their activities will never be able to control all the secondary effects, giving us a means for intelligence. On the other side of the same coin, any decision maker is responsible for charting a course to achieve their goals, and success is more likely if they can understand how the environment is changing, and perhaps just as important, how this change will evolve in response to their actions. Machine learning methods are able to model relationships in an environment, but commonly only those which can be directly measured with consistency. Tensor completion approaches provide ways of measuring hidden features, dealing with sparsity and errors, and accommodating the complex web of effects in an environment. Naively it can be thought of as building models where "everything predicts everything." These models can fill in sparsity, identify and correct anomalies, and make forecasts with data that would usually stymie other approaches. Still, tensor completion is non-trivial with few ready-made libraries available, the need to create tensor representations, and significant processing requirements which mandate distributed computing. Applying tensor completion models to simulate the impacts of scenarios (like future decisions) requires appropriate problem formulation which is non-obvious. NuWave has created the Artemis Anticipatory Network (A2N), an anticipatory intelligence platform which uses tensor completion to forecast behaviors, assess the impact of actions, track the likelihood of outcomes, explore discovered relationships, and detect anomalies. Built with a variety of stakeholders in the US Military, A2N allows users to deal with the variety and scale caused by the proliferation of modern sensors, providing interactive data exploration (including 60 period forecasts and alerts), and near real-time “what-if” analysis and scenario tracking. For the military, A2N provides the automation which accelerates analysis enabling users to effectively operate in the multiple domains and hidden actions of hybrid warfare. A data-driven, fully inductive data mining approach powering A2N helps mitigate human bias in assessments and improves whole-of-environment analysis. A2N is implemented as hybrid-cloud, ephemeral (serverless) microservices to ensure that it can scale on demand but has negligible persistent resources costs. This aligns with our customer’s event and scenario driven mission model characterized by periodic surges in activity to address new situations. When A2N's API is accessed resources are elastically provisioned, often processing many billions of records from multiple sources for multiple concurrent and/or serial tasks. Resources and workers provisioned to meet these needs disappear once their work is complete, although any data artifacts are retained, saving money and increasing peak processing speed over non-elastic approaches. Through an API-centric microservice approach, VANE also supports plugging in supplementary approaches for modeling, data acquisition, and other tasks rapidly and with no danger to existing operations.