Research News: Optimizing Sonobuoy Placement using Multiobjective Machine Learning
.jpg)
Christopher Taylor, Post Doc researcher at the Signal Processing Group, provides an overview of his research paper 'Optimizing Sonobuoy Placement using Multiobjective Machine Learning' which he'll present at SSPD Conference, 13/14 September 2022, IET, Savoy Place, London.
Summary
We have developed a new machine learning technique to finding optimal patterns for the placement of fields of sonobuoys in a complex undersea environment. Sonobuoys are portable, expendable sonar systems used for a wide range of applications including localization of underwater targets. The sonobuoys are placed sequentially in a pattern by an agent such as a helicopter, aircraft, or UAV. In our localization scenario we seek to minimize both sonobuoy placement time and uncertainty over target localization. Our approach allows an operator to choose between different optimal solutions, favouring lower placement time or lower localization uncertainty as operational circumstances require.
Importance of the research
In any localization problem, minimizing the uncertainty over localization is clearly a priority. Localization of underwater targets in a complex environment affected by various sources of noise and clutter is a difficult task, and sonobuoys are a key component. In a scenario where a field of sonobuoys must be placed to localize a moving and possibly stealthy target however, time may also be a critical factor, as shorter placement times may avoid the target moving out of range or taking some other action. Furthermore, shorter placement times may limit the number of sonobuoys that need to be placed, which may be important both in terms of economic cost and to avoid exceeding the payload capability of a placing aircraft. Our biobjective approach allows an operator to choose between different optimal solutions reflecting the trade-off between placement time and localization uncertainty.
We seek to improve on standard placement patterns, such as grids, lattices and ovals, according to operational circumstances. We use a hybrid machine learning approach, with two phases. In the first phase, a multiobjective evolutionary algorithm (MOEA) solves an offline problem, before the sonobuoys report new information, and produces a large population of solutions, assessed against the dual objectives of minimizing pattern placement time and target localization uncertainty. In the second phase, a multiobjective reinforcement learning (MORL) algorithm uses updated information from the sonobuoys to solve the online problem and further improve results.
We find that the MOEA improves significantly on standard grid patterns and that the MORL phase improves further on the evolutionary phase, at least with modest numbers of sonobuoys. The level of uncertainty over localization may also be reduced to a required level with a smaller number of sonobuoys.
What comes next?
We are looking to improve the performance of the algorithm further, both to handle more real-world complications such as drift and failure of the sonobuoys and if necessary, to generate optimal patterns using larger numbers of sonobuoys. This will involve use of more sophisticated computational and algorithmic techniques. We will look to use massively parallel computation, and to improve the algorithm in particular by developing new functions within the MORL phase for approximating the expected reward for placing the remaining sonobuoys in a given set of locations.
Link to Sensor Signal Processsing for Defence Conference.