SEED INNOVATIONS, LLC — Department of Defense STTR Phase I: MDA22-T004

Description

Seed Innovations (Seed) and the University of Colorado at Denver (CU Denver) apply our combined experience in artificial intelligence, machine learning, model interpretation and human-machine teaming to research and develop a prototype Deep Reinforcement Learning (DRL) explainability platform. This prototype solution, named ‘Reveal’, seeks to empower decision-makers by providing them with meaningful insight into DRL behavior. Through the research on this project, Seed and CU Denver explore methods to improve the explainability of both existing and new DRL models. Reveal analyzes existing models leveraging proven methods of manipulating model inputs and assessing model output sensitivities to those inputs. When new models are developed, Reveal offers a novel strategy of simultaneously training a DRL model while extracting interpretable rules made possible by exposing hidden layer activations from within the DRL model. Once the DRL model has been trained, the Reveal platform exports the derived rule sets and passes them to a rules engine capable of quickly applying the rules while simultaneously explaining which rules were applied and why. Whether working with an existing DRL model or a new model developed using the Reveal platform, post-processing occurs that utilizes natural language processing to translate identified sensitivities and rules into a human-interpretable narrative that provides value to all decision-makers, regardless of their experience or understanding of AI/ML. Approved for Public Release | 22-MDA-11339 (13 Dec 22)