全球智库信息集成服务系统

• The research team built models, based on a deep neural network architecture, that predict the readiness level of military units and organizations. The models did well at interpreting natural language descriptions of personnel, equipment, and training factors, as well as other extenuating information, in terms of readiness implications. The best model was able to correctly calculate which of four readiness levels a unit would report 75 percent of the time.
• Such a model could provide real-time feedback to unit commanders as they submit their monthly readiness reports to improve the accuracy and detail of those reports.
• General-purpose public domain word embeddings are good on many natural language processing (NLP) tasks, but such embeddings do not work as well in a domain, such as national defense, that has a specialized vocabulary and semantic context.
• On intermediate, task-specific word relatedness and analogy tests, defense-specific embeddings appear to significantly outperform public domain embeddings and would likely be useful in downstream NLP models dealing with defense-related matters.
• Multilayer neural networks as a whole performed much better than the single-layer logistic regression model used as a baseline; once a single recurrent layer was added, model performance further improved, but additional recurrent layers made little or no difference.

Military readiness is a perennial priority for the United States and a cornerstone of national security. Key to managing and improving readiness is the ability to measure it. This gives leaders situational awareness and tools for exploring trade-offs with other priorities, such as modernization, force structure, and use of national resources. There are likely many ways in which artificial intelligence (AI) can improve measurement and management of military readiness. In this report, the authors discuss work that advances the capability of computers to "understand" human language describing factors that promote or impede readiness.

The U.S. military reports monthly on overall readiness. These quantified reports are accompanied by narratives explaining what is occurring in military units that is affecting current or future readiness. The authors' goal in this report is to use these assessments to calculate overall readiness and enable senior leaders to estimate how readiness could be affected by personnel, equipment, or training factors. An additional benefit would be to have automated, real-time interaction with unit commanders as they write their assessments to help them refine the information they provide and better align their narratives with reported readiness levels.

• Chapter One

  Interpreting Military Unit Readiness Through Machine Learning

• Chapter Two

  Defense-Specific Word Embeddings

• Chapter Three

  A Deep Neural Network to Estimate Readiness

• Chapter Four

  Next Steps and Applications

• Appendix

  Word Embedding Primer