Complete final report
- Final bug squashing
- Drafted final report
- Completed final metric testing
- Reviewed and completed final report
The final outcome is of a reasonable quality. There are a large number of improvements that can be made but the base standard required has been met; a computational method to develop tactically sound plans based off enemy and friendly locations and a set of terrain data aligned with Geoscience Australia’s data.
Improvements (as per the final report):
- Investigation into optimisation of graph and cost field creation. This is currently the largest time sink and is purely a linear operation.
- Optimisation of candidate location population by use of Event callbacks in order to remove further additional linear operations.
- Improve the quality of the EA used. A likely intitial step is likely to be implementing roulette wheel parent selection in order to increase crossover diversity.
- Implement viewshed layering; keep base viewshed but maintain a `layered’ viewshed considering observation from other points (where the lowest viewshed value from all considered points is stored as the value for each location). This will allow for both complete enemy position LOS assessment and enemy observation post consideration.
- Implement additional cost check in preprocessing of candidate FUPs to ensure FUPs with no access via concealed routes are not considered in the candidate list.
- As a counter to the previous; quantitatively assess the merits of minimising preprocessing due to the speed of the EA.
- Implement timing check in preprocessing to determine if the total amount of candidate solutions is best assessed linearly (if the number is low enough to ensure the best plan is selected) or via EA (for large datasets as outlined in this paper). Note this is likely only feasible for terrain sizes of less than 1km square.
- Further development and optimisation of the fitness functions.
- Implement graph instead of list representation of candidate FUP/SBF locations in order to preserve spatial links and allow more effective mutation.
- Maintenance of a pareto front of non dominated solutions (based off FUP and SBF considerations) in order to provide a range of tactical courses of action.
- Undertake sensitivity analysis of fitness function in order to determine the effectiveness of the current implementation.
- Investigate implementation of tournament selection in the EA with an Artificial Neural Network (ANN) trained in comparing two plans as selector. With sufficient training, the ANN is likely to be effective as a fitness function as it is able to quantitatively identify the relationships between the inputs and the relative merits of a plan.
- Implement force element size requirements; currently the FUP and SBF locations are merely points (5m spacing). Adding a spatial considerations would allow effective up scaling to larger force elements.
- Include weapons systems employed in SBF. Fitness functions can then be modified from default based off the differing effective ranges of these weapons systems.
- Additional tactical considerations; timings, use of offensive support and fire planning, likely enemy withdrawal routes for cutoff position sighting etc.
- As this has been accepted for presentation at a Simulation conference, after the academic requirements are met, ongoing refinement will occur, in particular looking at improving the EA, widening the testing regime and, ideally, incorporating testing in a co-planning environment with impartial human planners.
Initial drafting – 4 hours
Metric testing – 3 hours
Final report – 8 hours
Total running time: 143 hours