Ongoing minor work. Investigate methods (Artificial Neural Networks (ANN)) that can optimise fitness function.
- Developed and rehearsed initial presentation
- Began development of report
- Developed objectives to optimise for multi objective optimisation
- Investigated ANN and began development of ANN from scratch
Report framework developed:
- Skeleton of full report
- Elements beginning to flush out (eg. Pathfinding options analysis)
Initial objectives developed:
- FUP1 : Concealed route
- FUP2 : Assault distance IN enemy LOS (accounts for undulations in assault)
- FUP3 : Total assault distance
- FUP4 : Elevation change in assault (Between FUP and Enemy postion OR Total cumulative elevation change – accounts for large undulations in assault)
- SBF1 : Observation to enemy position from < 0.3m
- SBF2 : Distance to enemy position as a distribution centred around 250m (Custom function likely or skewed normal distribution with sharp drop off at larger ranges and slower dropoff at closer distances)
- SBF3 : Elevation above enemy position
- PLAN : Angle between SBF and FUP – Exact function TBC however initial distribution function options below:
- Chi distribution centred on 90 degrees with n = 8 as per http://www.statisticshowto.com/probability-and-statistics/chi-square/#chisquaredist
- Alternate: Normal distribution with peak above 1 but capped at 1 (allowing a ‘flat’ distribution over the angle range (eg) 70 – 100 degrees.
No effective fitness function to collapse the multiple FUP and SBF objectives into a single score for each as yet (ie. weighting of individual scores)
Background understanding in Neural Networks (conceptual understanding). Work on Perceptron system being developed from scratch. So far a robust system has not been developed but the concepts are being solidified.
Presentation : 3 hours
Report : 3 hours
Objective development : 3 hours
ANN work : 6 hours
Total 15 hours (79 hours running total)
The initial presentation went well however the initial draft was 20 min+. This was an issue for a 7-8 minute presentation! The presentation was cut down to ~8 minutes however the military terminology background (eg. FUP/SBF definition, general tactical considerations etc) was one of the areas that was removed. The result of this was in effect a language barrier in some areas of the presentation between myself (using tactical words and concepts) and the audience who (largely) were not experienced in this area. The final presentation will need to incorporate some background and the intent will be to incorporate this into a brief (~30 second) planning process example in order to provide a good context to the audience.
It has been mentioned that any more than 3 objectives will reduce the effectiveness in EAs therefore the aim is to reduce the current 8 objectives to one score for FUP, one for SBF and one for the plan in total. This brings us back to the initial problem in that it is very difficult to effectively quantify a score for an FUP against another FUP (Especially in a generic sense which can be applied over varying terrains).
Investigation into ANN led to the realisation that if an ANN can be trained to identify the better plan out of two options (given the 8 objectives previously outlined as inputs for each plan) then the relative weightings for the neurons can tie into a weighting system for the fitness function.
This appears to be a relatively unique use for ANNs and I have not been able to find any reports of them being used in this way. The ANN would be trained by presenting training data consisting of a pair of plans as input with an output indicating if the first plan is a better plan (1) or worse plan (0) than the second. This data would require human input in identifying the better plan however the plans themselves could be randomly generated, thus the human interaction would merely be cycling through the (100s of?) plan pairs to confirm which is the better plan.