Scenario Development and Enhancement for Military Exercises
Navy SBIR 20.2 - Topic N202-126
Office of Naval Research (ONR) - Ms. Lore-Anne Ponirakis email@example.com
Opens: June 3, 2020 - Closes: July 2, 2020 (12:00 pm ET)
N202-126 TITLE: Scenario Development and Enhancement for Military Exercises
RT&L FOCUS AREA(S): Artificial Intelligence/ Machine Learning
TECHNOLOGY AREA(S): Information Systems, Human Systems
OBJECTIVE: Develop a user-friendly capability to create background information and long-form exercise injects derived from seeds, drawing from a catalogue of already extant background material for content.
DESCRIPTION: Information environments in situations of conflict and warfare are hectic, chaotic, and hard to predict. Military exercises and training capabilities currently lack realistic material to help them to develop Tactics, Techniques, and Procedures (TTPs) and blunt information attacks; and compete effectively in conflict situations. Warfighters require realistic training capabilities and the capability to develop, test and validate TTPs for information maneuver; this requires realistic, rapidly generated content to facilitate scenario development, enhancement and maintenance during an exercise or training experience.
Simulation of the information environment is a difficult problem. Simulating Facebook posts, blogs, and other long-form inputs (200 to 800 words) is labor-intensive and difficult to scale. Artificial intelligence breakthroughs have created new capabilities to generate realistic content that would be suitable to support an information environment simulation.
The GPT-2 model [Ref 1] and potentially other artificial intelligence solutions [Ref 2] provide useful starting points for realistic text simulation. This new language model and potentially other unsupervised multitask learners [Ref 3] have been demonstrated to perform downstream tasks in a “zero-shot” setting without any parameter or architecture modification.
The desired capability is the capacity to generate realistic information inputs for simulated training and exercise environments. The capability should be able to generate text to fit in multiple formats (Facebook, blog, other social media) - posts of 1-2 paragraphs, and long posts of 800 words (in English). The desired capability will have the ability to develop scenario materials for a refugee crisis, disaster scenario or a similar complex event; to edit and perform quality checks; to change and shift narratives; to add new events; and to catalog and index materials.
PHASE I: Define and develop an initial capability for generating 50 to 200 word (approximate) synthetically generated texts with a user interface to allow for review, editing, tagging, and flagging of the produced material for initial assessment. Produced texts should be packaged to enable the flow of the materials into databases for input into a synthetic environment reservoir for test and evaluation. Develop a Phase II plan. Phase I Option, if exercised, will expand materials to develop a catalogue of synthetic background data, improve the user interface, and institute a tagging, flagging, and automatic or semi-automatic indexing function.
PHASE II: Develop a data editor and visualization capability to assist White Cell scenario authors to create narratives, inject new discourses and gists, and review the gists, discourse material and narratives in the reservoir. Improve the fidelity and capability of the Phase I product to generate texts from background materials, created by scenario authors, that are sufficiently realistic and on topic to meet a minimum of realistic level of volume and velocity (10K tweets an hour, 100 longer (up to 800 word) posts per hour).
PHASE III DUAL USE APPLICATIONS: Develop the capability to attach discourses to personas and adjust texts to conform to target narratives and discourses, so that scenario creators can develop realistic stories for military exercises and training. Demonstrate the capability of flowing synthetic texts into simulation technologies and tools so that they can be used in an information conflict war game scenario, inject new material, and provide a realistic volume and velocity of data for a training exercise (50K tweets/hour, 1000 posts/hour). Investigate the feasibility of the capability to synthetize texts in other languages, to answer questions, and to perform translations.
Marketing and brand name companies also require new capabilities to train staff in information conflict to support their brands when dealing with trolling, meme conflicts, and other social cyber-attacks in the information environment. Non-profits such as the Red Cross and other Western aid agencies have problems similar to the U.S. Government in defending their message against foreign attackers, seeking to diminish their reputations among target audiences.
1. “Better Language Models and Their Implications.” OpenAI, February 14, 2019. https://openai.com/blog/better-language-models/
2. Vig, Jesse. “OpenAI GPT-2: Understanding Language Generation Through Visualization.” Medium, 5 March 2019. https://towardsdatascience.com/openai-gpt-2-understanding-language-generation-through-visualization-8252f683b2f8?gi=fc3e151fc89f
3. Radford, Alec, We, Jeffrey, Child, Rewon, Luan, David, Amodel, Darlo and Sutskever. Ily. “Language Models are Unsupervised Multitask Learners.” https://d4mucfpksywv.cloudfront.net/better-language-models/language_models_are_unsupervised_multitask_learners.pdf
KEYWORDS: National Language Processing, Artificial Intelligence, Information Operations, Military Exercises, Training
TPOC-1: Rebecca Goolsby
TPOC-2: Amy Bolton