SCRO / Journées de l'optimisation

HEC Montréal, 29-31 mai 2023


HEC Montréal, 29 — 31 mai 2023

Horaire Auteurs Mon horaire

MDSAII Military, Defense, and Security Applications II

31 mai 2023 15h40 – 17h20

Salle: Procter & Gamble (vert)

Présidée par Francois-Alex Bourque

4 présentations

  • 15h40 - 16h05

    Naval Fleet Schedule Optimization using Mixed Integer Linear Program

    • Widmer Megan, prés., DRDC
    • Michele Fee, DRDC-CORA

    When it comes to future fleet planning, having a way to quickly model the readiness of ships over time is a useful decision support tool. Optimizing fleet schedules provides a mapping of each ship’s readiness according to a set of fleet-wide requirements. A single ship’s schedule can be modeled as repeating, idealized operations and maintenance cycles that track a ship’s readiness over time. The fleet schedules can then be produced from offsetting these from ship-to-ship to meet a fleet’s readiness requirements. By expressing these requirements as a series of linear constraints, it is possible to optimize these offsets using a mixed integer linear program with an off-the-shelf python solver. To verify the approach, a series of optimized fleet schedules were generated for notional platform fleets of different sizes. The schedules produced were compared qualitatively to those generated with an existing genetic algorithm based fleet schedule optimizer. Between the two tools, the linear program was found to better meet the schedule requirements.

  • 16h05 - 16h30

    Deconflicting Helicopter and Unmanned Aircraft System Operations at Sea Using Mixed-Integer Linear Programming

    • Jean-Denis Caron, prés., Defence Research Development Canada
    • Michele Fee, DRDC-CORA
    • Trevor Phair,

    The Centre for Operational Research and Analysis (CORA) of Defence Research and Development Canada is responsible to provide operational research and analysis support to the Department of National Defence and the Canadian Armed Forces. In 2021, CORA conducted a study to help the Royal Canadian Navy with a project related to the acquisition of an Unmanned Aircraft System (UAS). More specifically, the objective was to de-conflict the Maritime Helicopter (MH) and UAS operations at sea by generating flight schedules optimized to provide maximum Time-On-Station while taking into account onboard space constraints. A Mixed-Integer Linear Programming (MILP) model was created and used in conjunction with a Monte Carlo simulation, which accounted for the platform serviceability (MH and UAS). The main drivers included the maintenance cycle of the MHs and the UASs, the crew rotation on-board the ship and deck/hanger limitations. In total, more than 40 unique Task Group (TG) configurations were considered. Each TG was composed of up to four ships, up to two MHs with up to two crew each, and up to three UASs. This presentation provides background and context on the problem, discusses the main assumptions, provides an overview of the MILP and introduces some of the results.

  • 16h30 - 16h55

    Using POMDP reinforcement learning to find approximately optimal submarine search patterns

    • Alexander Milne, prés., DRDC
    • Pelletier Emile, DRDC

    We examine the use of a partially observable Markov decision process (POMDP) to model and find solutions to a submarine sanitization problem. Submarine sanitization is a process where an area is searched by a submarine to determine whether there is another underwater vehicle present. The agent is constrained in terms of the range and probability of detection, and its ability to maneuver. As well, detections may be false positives. A POMDP is a stochastic process that consists of action, observation, state and reward elements. Its solution determines the sequence of actions to maximize long-term reward. Rewards depend directly on the actions and states they are taken in, whereas the agent chooses actions based on belief states informed by observations. A formulation of the submarine sanitization problem into a POMDP will be discussed, as will the application of POMDP solvers to find approximate solutions. Using an open-source solver, we obtain approximately optimal action choices which maneuver the agent to areas with high probability of target presence. A change in parameters yields a change in behaviour: the agent requires a higher certainty to call off the search when given an increased incorrect guess penalty. Increasing state and observation complexity reduces sanitization performance.

  • 16h55 - 17h20

    Imbalanced data in marine platform activity classification using equipment health monitoring signals

    • Scott Koshman, prés., University of Toronto, Canadian Forces College
    • Fae Azhari, University of Toronto

    Marine platforms and their machinery will operate in different contexts and states throughout their life and under varying duty cycles. In the development of reliability and maintenance models for this machinery, inclusion of context may improve fidelity and utility for decision making processes. For partially available context information, a method of compensating is to substitute these gaps with context classifiers trained from the ship’s equipment health monitoring data records. In other words, the feedback signals from the equipment are used to inform the analysis by reverse engineering the equipment’s operation. A challenge noted while developing these classifiers is that the field data, from which the training data is extracted, often exhibits imbalances in class distribution, which can create negative bias in supervised learning away from the minority class(es). Utilizing the online equipment health monitoring for classification of macro ship activity, approaches to mitigating the impacts of imbalanced data are investigated.