Research Topics

Below is a list of research topics supported by the AFRL. Use the filters and keyword search below to find research topics of interest. You can apply for up to 3 topics on your application.




Scholars are encouraged to contact any mentors whose projects they find of interest. To contact the mentor, use the link included at the conclusion of each project description.

Characterizing Swarm Systems Capabilities
Mentor: Jeffrey Hudack, Information
Location: Rome
Academic Level: Masters, Ph.D., Upper-level Undergraduate

Large collectives of coordinating robotic systems require new methods for characterizing expected behavior and communicating that capability to human users. Interactions between platform capability, hardware configuration, and software algorithms yield a mean and variance of expected performance for a given mission. Characterizing these interactions and resulting performance requires new analytical methods that can scale to hundreds or thousands of entities and can map a manifold of performance to specific mission objectives. We seek new methods for characterizing expected swarm performance, with tools that can help a human operator to explore the trade-off space of number of platforms, configuration, and software algorithms. Areas of interest include Operations Research, Statistical Analysis, Multi-Robot Task Assignment, Modeling and Simulation, and Dynamical Systems.


Data Efficient Machine Learning
Mentor: Walter D Bennette, Information
Location: Rome
Academic Level: Masters, Ph.D., Upper-level Undergraduate

A trained classification model can be used to automatically sort items of interest, such as images or signals, into discrete categories or classes. Modern machine learning algorithms typically require a vast amount of labeled data with which to learn such a classification scheme. Unfortunately, this quantity of data is sometimes difficult to acquire. We are interested in techniques that help users efficiently label datasets to reduce the amount of labeled data required to achieve a useful classification model. Your work will focus on implementing state of the art techniques for data efficient machine learning, running simulation experiments, and helping develop the next generation of efficient labeling procedures. Other projects will be introduced over the summer with opportunities for additional involvement.


Generalizing from Few Examples with Meta Machine Learning
Mentor: Matthew M Klawonn, Information
Location: Rome
Academic Level: Masters, Ph.D.

State of the art machine learning models have achieved good generalization performance in a range of application domains, such as computer vision, natural language processing, etc. This is surprising because these models often contain an order of magnitude more trainable parameters than there are available data points. According to traditional ML wisdom, such models should overfit the training data. This topic proposes to investigate emerging theories explaining why overfitting does NOT occur in such overparameterized models. In doing so, we hope to uncover insights that will allow for learning in domains with highly limited training data. Our intuition is that meta-learning, i.e learning the learning process, will play a role in a data efficient learning regime.


Interactive Machine Learning
Mentor: Matthew M Klawonn, Information
Location: Rome
Academic Level: Masters, Ph.D., Upper-level Undergraduate

Many recent efforts in machine learning have focused on learning from massive amounts of data resulting in large advancements in machine learning capabilities and applications. However, in certain domains obtaining explicit labels for training data can be challenging. In such cases, it is often desirable to collect feedback from subject matter experts via queries that elicit implicit information. For example, when ranking items as is done in search algorithms, it is often easier to ask a user to compare two items than to assign each a number in isolation, and rank them later. Posing such queries and learning from such feedback falls into a subfield of machine learning called interactive learning. Scholars will be asked, with help, to read and understand materials describing existing interactive learning techniques, and apply them to novel domains that may include, multi-objective optimization problems, , generative models, and more.


Machine Learning for Autonomous Vehicle Control
Mentor: Ashley Prater-Bennette, Information
Location: Rome
Academic Level: Masters, Ph.D., Lower-level Undergraduate, Upper-level Undergraduate

This topic will demonstrate the use of machine learning models in training and deploying autonomous agents on small platforms. The agents will be developed and trained in a simulation environment using reinforcement learning and other machine learning techniques. After thorough software simulations, the agents will freely navigate small platforms using only the data available from its onboard sensors to navigate a complex environment and achieve a specified goal. The demonstration vehicles may include small robotic cars (such as Lego Mindstorm kits) or quad-copter style drones.


Machine Learning for Multi-Domain Operations
Mentor: John E Myers, Information
Location: Rome
Academic Level: Masters, Ph.D., Lower-level Undergraduate, Upper-level Undergraduate

Design and develop machine learning models for the processing and classification of data associated with operator interactions within a Multi-Domain Operations Center (MDOC). Work will include using commercial tools to process voice chat data, the application of ontologies to semantically link the chat information to operators and specific missions, and machine learning models for auto-responses to inquiries.
An example of a Concept of Operations (CONOP) would as follows: the Senior Offensive Duty Officer (SODO) speaks through their head set “What is the ETA of POOH71?” Using speech-to-text libraries, extract which audio line spoke (SODO) and the sentence spoken, and run the text through a trained model. This model will be trained to classify the sentence based upon a variety of categories given. Once the sentence is properly classified as an “ETA_Request,” do a natural language extraction of the noun entity, in this case “POOH71.” Once the classification and noun are deduced, perform a web service call into Codex to return the ETA of that aircraft call sign. Finally, in response to the SODO’s initial question, the computer will verbally respond to the query with the ETA of the aircraft.


Mathematical Theory for Advances in Machine Learning
Mentor: Ashley Prater-Bennette, Information
Location: Rome
Academic Level: Masters, Ph.D., Upper-level Undergraduate

As machine learning methods are being increasingly adopted into decision-making technologies, it is essential that the performance of the machine learning techniques continue to improve. This research topic is focused on applying recently developed mathematical techniques to machine learning and pattern recognition methods in order to improve accuracy, interpretability, data efficiency, and robustness. Special emphasis will be placed on techniques that admit low-rank or sparse representations of multidimensional data.


ML-Boosted Knowledge Graphs
Mentor: Nicholas Ricky Del Rio, Information
Location: Rome
Academic Level: Masters, Ph.D., Lower-level Undergraduate, Upper-level Undergraduate

This summer program will focus on using the Semantic Web to streamline the discovery, assembly, and consumption of bespoke machine learning (ML) datasets in transparent and reproducible ways. Specific challenges include how to obtain a knowledge graph of ML training data simply as a byproduct of usage. Additionally, this program will investigate next-generation (SoS) architectures that connect symbolic reasoning systems with statistical ML knowledge to support high-quality question/answer scenarios. Specific challenges include context-aware graph embedding algorithms and the inverse—how to propagate derived results in metric spaces back to symbolic graph space.


Operations Research Supporting Command and Control Decisions
Mentor: David Myers, Information
Location: Rome
Academic Level: Masters, Ph.D., Upper-level Undergraduate

Operations research deals with the application of advanced analytical methods to help make better decisions. While data-driven decision making has a long history in the planning processes of the Air Force, as newer warfighting domains begin playing more prominent roles in AF operations, the need for data-driven decision-support tools is growing rapidly. This project will explore various aspects of the application of operations research techniques to the problem of quantifying courses of action in all warfighting domains. It can include the research and use of design of experiments, machine learning, graph theory, multi-criteria decision making, as well as constrained optimization.


Planning for Wargaming
Mentor: Brayden Daniel Hollis, Information
Location: Rome
Academic Level: Masters, Ph.D., Upper-level Undergraduate

This topic seeks to develop automated planners that can intelligently achieve missions in adversarial environments. Wargames have long been an effective method for developing and validating tactics, but traditionally rely on human players to provide sufficient challenge and variety. Students’ work may include encoding wargames into machine readable formats, implementing existing automated planning techniques to develop intelligent A.I. opponents, and developing new approaches for planning against an adversary. Possible extensions also include interactive planning techniques that allow automated planners to work with humans to develop robust plans.


Reinforcement Learning and Planning for Trustworthy Autonomy
Mentor: Alvaro Velasquez, Information
Location: Rome
Academic Level: Masters, Ph.D.

Reinforcement learning and sequential decision-making have been revolutionized in recent years thanks to advancements in deep neural networks. One of the most recent breakthroughs was accomplished by the AlphaGo system and its victory over the world Go champion. However, even in this impressive system, the learned agent performed sub-optimal actions that puzzled both the Go and the reinforcement learning communities. Such failures in decision-making motivate the need for methods that can provide (statistical) guarantees on the actions performed by an agent. Under this topic, we will explore novel decision-making techniques from the model-free perspective of reinforcement learning as well as from model-based views of control theory and planning. As a second technical aim, we will integrate statistical guarantees on the behavior of our learned agents. This is an open-ended research endeavor that seeks to provide a minimally intrusive direction on the students’ own intellectual proclivities. Since reinforcement learning and decision-making are highly interdisciplinary topics, the idea is for candidates to leverage their own knowledge and interests on this set of highly relevant problems.


Wargame AI Development (WAID)
Mentor: Ainoghena Igetei, Information
Location: Rome
Academic Level: Masters, Ph.D., Upper-level Undergraduate

Advances in machine learning and artificial intelligence (AI) has enabled breakthroughs in developing intelligent agents capable of beating human experts in complex strategy games. The space and action complexity of some of these games (i.e. StarCraft) closely model scenarios relevant to Battle Management Command and Control (BMC2), where we operate in large decision spaces, variable timelines, and most importantly we are constantly working in states of imperfect information. This project looks to leverage the advances in machine learning and AI to develop intelligent agents capable of producing high level strategies. We envision that these agents will serve as an aid to develop new battle plans and also serve as intelligent adversaries to evaluate the success rate of existing plans.