Key Points of Understanding

Setting up the optimization simulation is critical to the final outcome. Important aspects such as choosing plausible locations and orientations for the detectors, risk grades of zones and sub-zones highly influence the time for calculation and quality of result.

Just as important is choosing the detector layout once the optimization completes, as several different layouts are typically available to choose from. Once the optimized layout is added, it can also be modified as a normal layout in Detect3D – you are not restricted or required to use the optimized layout!

As the user, you are in complete control – the algorithm is simply performing the number crunching.

A common misunderstanding is that the final result of the GA is the one optimal detector layout that is better than any other possible layout. Unfortunately, the nature of a global optimization search is that we can never know this – there are simply too many possible layouts to test each one, so it is always possible that there is a better layout than the one produced by the optimization. It is even possible that the GA could find a better layout by doing the same simulation twice, as it is a stochastic process.

From experience, there are many optimal layouts for any given simulation setup. The layouts are typically grouped in “families”. The optimization will virtually always arrive at one of these families, but which one, and whether it’s “the optimal” family, can never be known.

The correct way of thinking about the results of the optimization is not that it provides “the optimal layout” as we can never know if this is true. Rather, the optimization provides “an optimized layout”.

A very common misconception is that an individual is a detector. It is not - an individual is a layout made up of several detectors. Understanding this point makes the setup of the simulation much clearer.

The Genetic Algorithm is a stochastic process – given the exact same setup, it is highly likely that the results will be different layouts, with slightly different coverages. It can sometimes make sense to run the same simulation more than once as a quality check.

The algorithm will not achieve anything more than what you ask of it. For example, if you set the target for a layout to be 90% 1ooN, then it will be happy with a layout that is 90.01% 1ooN, and not judge a layout that is 99% 1ooN to be better, since both are over the target.

If you want 99% 1ooN coverage, then set the target to 99% 1ooN.

While this may be possible in very simple cases, the most likely explanation is that you have given yourself a target that you did not ask the algorithm to achieve.

For example, if you have set up the algorithm to automate and minimize detector count, and it has found a layout with 3 detectors and 91% 1ooN coverage, then it will assess this as adequate and will not strive to achieve 95% 1ooN with 3 detectors coverage. Instead of improving the coverage, all its efforts will be focused on attempting to reduce the detector count to 2.

Make sure that you have given the optimization the same goal as you have yourself!

The algorithm works best the more “hands off” you are. For example, if you have an area that you believe can adequately coverage with 6 detectors, then it is preferable to initialize the detector count to 2 detectors and allow the algorithm to automate the detector count – it will arrive at 6 detectors if that is indeed the right number. The alternative of setting 6 detectors and not allowing the algorithm to add or remove detectors may also result in an adequate layout, but it is a less reliable method. The optimization may also surprise you by finding a layout with fewer than 6 detectors.

If the detector count is automated, the algorithm will often add more detectors than you believe should be in the final layout. This is normal and is expected– it is learning from these layouts and will remove detectors in future generations.

A feature of any Genetic Algorithm is that there is no well-defined “end-point”. While there is a setting to stop when a certain number of generations is reached (the default value is 250) it is likely that most of your simulations can be stopped before then. The most important things to look for as indications that the Genetic Algorithm cannot improve on the current layout are outlined below.

If you have set the algorithm to automate the detector count, you should stop the algorithm when:

1. The maximum fitness does not increase (i.e. it “flat-lines”) for 30 generations or more. It is important that you do not consider the best coverage flat-lines as an indicator that the genetic algorithm cannot improve – in many cases, the coverage will flat-line well before the fitness flat-lines, meaning that even though the algorithm cannot improve the coverage, it is finding a way to remove a detector. This is a critical point to remember, especially in complicated simulations where the coverage can flat-line for over 100 generations while the maximum fitness still increases.

    

   

   The above image indicates a section of the maximum fitness which has flat-lined for 30+ generations.

    

2. The coverage target has been achieved and the maximum fitness is increasing very slowly but is less than 0.7. To remove a detector, the maximum fitness must be about 1.0, and if it does not look like it will achieve a value close to this, it is a good indicator that the minimum number of detectors has been reached. The value of allowing the optimization to run longer is that the best layouts may be populated with more options to choose from.

If the detector count is fixed, you should stop the simulation when:

The best coverage does not increase for 30 generations or more. There is no need to consider maximum fitness in this case, as the maximum fitness and best coverage are synchronized such that they will both flat-line at the same time.