|
Fritz Jooste Administrator Posts: 81
5/26/2020
|
Fritz JoosteAdministrator Posts: 81
JunoViewer offers wide range of deterioration model types, as shown by the figure below. The model types fall broadly in two classes: (a) Ranking Models; and (b) Benefit-Cost Analysis (BCA) Models. Within each main model class, there are several options or sub-types that can be used. For example, within BCA models, you can choose whether to use an NPV-optimized approach or a Benefit Cost Ratio - optimized approach. With all these options available, clients often ask: "Which model is best?". This post deals with that question.
It should be recognized at the outset that there are many different objectives to an asset deterioration modelling task. For example, consider (a) an asset engineer looking to develop a Forward Works Plan (FWP) for the next four years under a known and fixed budget; next, consider (b) an analyst looking to determine the required budget to keep an asset network condition stable over a 30 year period.
Clearly, the objectives for these two modellers will differ. For person (a), it will be important to perform maintenance on as many segments as possible and to maximize utilization of the budget. For person (b) it will be important to determine an optimal budget and determine a treatment strategy that will optimize long term benefits to the network.
Furthermore, it is important to realize that even within a relatively narrow modelling task, there may be several objectives that need to be maximized. Sometimes, these objectives may compete. This is illustrated in the figure below, which depicts the outcomes of several different model runs, each of which uses a different approach (e.g. ranking or BCA) and decision strategies (e.g. different treatment types and trigger conditions). There are two objectives the modeller would like to satisfy - (a) maximize total network benefit; and (b) treat as many segments as possible over the modelling period.
Now, in the above example, which model is best? Clearly, model C is not optimal in either of the two objectives (not Pareto optimal). However, choosing between model A and B is more difficult. Model B achieves a higher benefit, but it treats fewer segments over the modelling period. This may happen if a certain model approach favors fewer projects with high cost and high benefit. These few large projects maximize benefit but consume a significant portion of the available budget. By contrast, model A strikes a balance between maximizing benefit and treating a high number of segments over the modelling period.
So the first task in modelling should be to perform an honest assessment of the real modelling objectives. We use the words "honest" and "real" here because in practice you may often find that superficial attention is paid to "maximizing benefit to the public" and so forth. However, when a practical, achievable operational plan is needed, engineers will tend to focus objectives that are more real and pressing, such as ensuring that the available budget is fully utilized, that a certain percentage of the network is sealed in each year, etc.
Unfortunately there is no way to know in advance which model type and rule set will best satisfy all of your modelling objectives. In optimization and machine learning research, this is called the "No Free Lunch Theorem". In essence, this theorem implies that if you make no assumptions about the data, then there is no definite reason to prefer one model over another. Thus for some networks and network conditions, one model may perform better than another, and for another network or condition, vice versa.
The conclusion is thus (a) understand clearly what your modelling objectives are; and (b) test and explore different models to find the one that best suit your model objectives for a specific modelling task; (c) know that next time round, a different model may be a better one to use!
edited by admin on 5/26/2020
|