Introduction and Background
This application analyses OPTAIN optimisation outputs and shall support decision making. While all solutions provided by the SWAT+ / COMOLA workflow are pareto-optimal (none of the objectives can be improved without losses in other objectives), choosing among a large number of solutions can be daunting.
To reduce complexity while minimising information loss, this application provides two ways to filter/reduce the pareto front:
- A clustering algorithm based on a Principal Component Analysis (PCA) and kmeans/kmedoids. The user can modify the clustering process, alter the number of tested clusters and the way outliers are handled or how much correlation is accepted across the considered variables. Finally, those optima representative for different clusters can be plotted and the measure implementation they recommend can be compared.
- An Analytical Hierarchy Process that can be run as standalone method as well as as additional feature on top of the clustered pareto front.
The application is structured the following way:
The second tab Visualising the Pareto Front provides an overview over the optimisation results. The user can gain insights into the relationships between the objectives and the pareto front by selecting and plotting preferred objective ranges.
The third tab Data Preparation is needed to produce the data required for the subsequent analyses. Several files need to be provided so the variables considered in the clustering can be calculated.
The fourth tab Configure Clustering allows to perform the clustering with default settings or to jump to the optional tabs for manual clustering.
- The tab Clustering Part 1 - Correlation Analysis can only be accessed if manual clustering is chosen in the Configure Clustering tab. It allows to assess and alter variables considered in the subsequent clustering.
- The tab Clustering Part 2 - PCA & kmeans/kmedoids provides the possibility to adapt, modify and finally perform the clustering process.
The Cluster Analysis tab lets the user plot the optima remaining after the clustering. Each of these optima is representative for one cluster.
The tab AHP - Analytical Hierarchy Process allows to determine priorities across the pareto front in a different way through assigning weights across the optima. It is possible to combine the clustering results with the AHP.
To ensure compatibility with algorithms (e.g. CoMOLA) designed for maximisation, some projects used negative numbers. Please note that, unless an objective uses mixed signs, this app omits the minus sign of these values. The interpretation however remains unchanged.
Data Preparation
This tab requires you to provide the optimisation outputs. Please refer to the Readme for examples of their structure. You can provide a limited set of outputs to only analyse the Pareto front in the next tab.
Alternatively, you can upload more data and prepare the variables for the subsequent correlation and cluster analysis by clicking Check Files and (if all files have been found) Run Prep. Please be aware that the preparation might take up to 15 minutes.
For being able to plot and analyse the Pareto front, please provide pareto_fitness.txt as well as the objective names. If you would like to plot the status quo, sq_fitness.txt is also required:
If you also want to run the subsequent correlation and cluster analysis, please provide the following files. Their names have to align with what is given here:
Run Preparation Script when ready (depending on the size of the shapefiles this can take up to 10 minutes)
Visualising the Optimisation Output
This tab plots the pareto front in a few different ways and lets you explore the effects of different objective ranges. You can select specific points/optima on the pareto front by clicking on them, then you can plot and download the map of the respective NSWRM plan. You can also select optima in the line plot and analyse their location in the objective space.
Range (slider selection)
Optimum (selected in Pareto plot)
Configure Cluster Settings
Clustering Part 1 - Correlation Analysis
A correlation analysis is needed as correlation among variables can skew cluster results. Therefore, please click Run Correlation Analysis. Based on the levels of correlation you can select those variables you would like to exclude from the subsequent clustering. Select them and then click Confirm Selection. You can come back to this tab to change the selection of variables later. It is also possible to run the clustering across all variables and select no variables to exclude in this tab, however please always click Confirm Selection.
Clustering Part 2 - PCA & kmeans/kmedoids
This tab requires you to decide on the cluster settings. After selecting how the objectives shall be plotted, deciding on the axis titles and confirming the number of PCAs tested, the clustering can be run with default settings. Selecting Yes under either 2. or 3. allows to change those default settings and test a variable number of clusters and outlier considerations.
The cluster outputs open in separate tabs and can be saved as images.
3.1 Please specify the number of standard deviations that shall be tested:
3.2 Please specify how many extreme variables within a datapoint shall be tested:
3.3 Please select a limit for the number of extreme solutions allowed in the final clustering:
Please specify how many clusters to iterate through:
Analysing the remaining optima
This tab allows you to analyse the cluster outputs and plot and compare the measure implementation across the pareto solutions selected in the clustering. The table shows those optima selected as representative for the different clusters. The plot on the right aligns with the one produced during the clustering. It shows the location of the optima selected in the table. Please be aware that plotting the measure allocation takes around 20 seconds.
Analytical Hierarchy Process
This tab allows you to run a different approach (AHP) to selecting pareto optima that best match your preferences. AHP is a decision making tool that helps you prioritise different objectives by comparing them in pairs. If you want you can limit the objective ranges and number of measures under 1.
Under 2. you can compare objectives two at a time and decide which objective is more important and by how much. ParetoPick-R will assign weights to each objective based on your inputs and check its consistency. The respective best choice is plotted below and you can decide whether it should be selected from the whole pareto front or from the subset of cluster results.