Skip to main content

Systems Learning Applied to the Tactical Environment (SLATE)

Systems Learning involves understanding and gathering information about systems from small organizations (such as individual ships or other unit level groups) all the way up to global entities (such Fleets or Nations). Systems learning can be placed alongside the term organizational learning and has been more evident in recent knowledge management research. As the proliferation of unmanned systems in the ocean and atmosphere continue to proliferate, so too do the degrees of freedom in the meteorological and oceanographic conditions needed to properly operate each individual system. Unmanned craft come in widely varying different sizes, weights, and power requirements (SWaP) with different sensing capabilities. These vehicles will have widely varying operating conditions that will tolerate environmental conditions differently.

This ONR Technical Candidate program will develop Systems or Machine Learning based correlations between the current weather conditions and pattern of life for ships and aircraft in a region. Scientific efforts under this thrust are aimed at leveraging existing tools, algorithms, and datasets to enhance understanding of environmental tolerance. This includes improved data and metadata creation and curation, systems learning including Bayesian data treatment and concept drift identification, and algorithmic outputs that develop empirical thresholds of operation and show movement of assumptions from the prior to posterior distribution.


Objective

The goal of SLATE is to blend information such as the physical ability (based on known safety or performance limits by platform class) for an asset to take-off or sortie from a specific port or airfield, with other trends for transit and on station behavior, and the effects of severe weather or other environmental boundary conditions on platform and sensor behavior, such as from effects of clouds, thunderstorms, and/or dynamic and thermodynamic meteorology. Investigators should identify algorithms that can incorporate proposed datasets with mixed data modes including binary, categorical, and continuous values.

Research thrusts for this program include, but are not limited to:

  • Identification and mitigation of concept drift: learning techniques to identify non-stationary aspects of the data and/or shifts in tolerance criteria
  • Iterative Bayesian updating techniques to nudge a model with new data on scene without a full re-training.
  • Software infrastructure development capable of full model retraining such that model updates minimize implementation of new information (e.g. new coefficients).
  • Modular algorithmic construction and infrastructure that supports reusable methodology to apply environmental, configuration, and pattern-of-life information generically to many different specific use cases.

Some example use cases for decisions from algorithm outputs may include:

  • What is the probability of being able to perform a small quadcopter launch test in 24 h, given uncertainty in the forecast environment and hardware tolerance?
  • Knowing we could not test our unmanned underwater vehicle yesterday under specific wind/wave conditions, how can I update my model on site to take this new data into account?
  • If I plan to fly a specific airline to reach a field site, can I determine that certain times of day or shifts have different tolerances for environmental delays or hardware breakdown tolerance?

Projects should indicate proposed meteorology and oceanography datasets to be used and the assumptions made in interpreting them compared to operationally available information. Particularly, it is expected that analysis/reanalysis data and/or current operational numerical weather prediction (NWP) output will be the main environmental data source. Sensitivity testing of data quality compared to multiple or degraded datasets are outside the scope of this project, but will likely be addressed in a future program.

Validation and verification (V&V) techniques should also be discussed in proposed work. Particularly, it is expected V&V will be developed alongside algorithmic development to support baseline skill metrics and improvements with algorithmic complexity. As the machine learning applications mature in sophistication and additional data is gathered, performance will be measured against previous model accuracy and adjusted Bayesian priors. These can be shown via confusion boxes, reliability diagrams, Briar Skill Scores, and/or other appropriate metrics.


Request for Planning Letters

The first step in the proposal process is for prospective investigators to prepare planning, allowing investigators to submit a short (three pages maximum) summary of their ideas on this topic for ONR to evaluate, provide technical feedback and indicate whether a full proposal would have a reasonable chance of success.

Refer to the current Marine Meteorology Planning Letter guidelines and indicate your desire to work with the SLATE project and team. Please note "SLATE Planning Letter ‘Your Last Name’" in your email subject line. If you do not receive acknowledged receipt within 10 days, please follow-up with a resend.

Important Dates

ONR funding cycles operate on the federal fiscal year. Funding amounts and timing may vary year-to-year. Please use the following dates as guidelines for full consideration of funding into the next calendar year.

August 15, 2023: Last date to submit planning letters (please submit by e-mail to below).

September 1, 2023: Estimated response deadline by ONR to all submitted planning letters with proposal recommendation.

October 15, 2023: Last date for proposal submission where the evaluation will still be eligible for full consideration in the FY24 funding.

All planning letters should be submitted by email to Kate Mulreany (katherine.l.mulreany.civ@us.navy.mil), Josh Cossuth (joshua.h.cossuth.civ@us.navy.mil), and Dan Eleuterio (daniel.p.eleuterio.civ@us.navy.mil).