Darwin Calibrator employs a powerful competent genetic algorithm search method based on the principles of natural evolution and biological reproduction. This kind of search algorithm is well suited to optimization of problems of a non-convex and multiple local-optimal solution nature. Calibration of a hydraulic model falls into this problem category and, as a result, a GA-optimization based search tool, such as Darwin Calibrator, is a sound choice for hydraulic model calibration.

Despite all the good features of GA there are, however, some issues to consider:

• A solution is fitter only in relation to other known solutions and, consequently, a GA has no test for true optimality. As a GA only knows the best solution relative to others, a GA has no precise rule for when to stop. This means that heuristic methods must be used to determine whether to stop a GA run. In Darwin Calibrator you can set a GA run to stop either by:
• Clicking Stop.
• Setting a maximum number of trial solutions.
• Setting a maximum number of non-improvement generations, whereby if the fitness of the best solution does not improve by more than a specified tolerance in a set number of generations, then the GA stops.
• A GA is a non-deterministic method that relies to a certain extent on its initial random population (starting locations in the solution space). Thus, each GA run performed may produce different solutions. (If you keep all GA parameters and fitness settings the same, the method is deterministic and will produce identical solutions every time.) Given the fact that a GA has no true test for optimality, after stopping a GA and producing a particular result, there is always the possibility that if you run the GA again you may find a better solution. In fact, it is good practice to run a GA a number of times, each time modifying something about the GA run (e.g., GA parameters, fitness weightiness, or adjustment group settings), in order to produce another set of potentially better results. At a minimum, the random number seed should be changed for each individual run so that the GA search initiates differently and therefore concludes differently.
• The GA calculates fitness of each trial solution according to the defined objectives for the optimization problem. GA only uses objective means to decide what constitutes a fit solution and what constitutes a less fit solution. The GA has no way of subjectively assessing a solution other than the methods (weightings) built into the definition of the fitness calculation. The best solution found by a GA shouldn’t be blindly accepted as being correct. To any single optimization problem there are likely to be many solutions that closely match the required objectives. Due to the fact that the GA has no concept of what constitutes a fit solution, other than its performance against the defined objectives, the GA may produce solutions that are impractical. That is, the GA cannot think for the engineer, it can only search the combination of choices that are presented to it. If the engineer doesn’t provide the GA with high quality data and enough or sufficiently flexible options to consider, then the GA may not be able to find a satisfactory solution. Conversely if the GA is presented with too many possibilities to try (e.g., in Darwin Calibrator, if you define excessively large adjustment group ranges combined with small adjustment increments and a large number of adjustment groups), then the efficiency of the GA search is reduced, and the likelihood that the GA will find the correct answer is also greatly reduced. GA is a highly sophisticated search technique, but despite all of its great features, GA still must be used with a degree of engineering judgment and skill. Only then can the engineer expect the GA to find solutions that are not only fit but are practical and likely to represent the real life situation as accurately as possible.
• Uncertainty in field observations should be assessed before these observations are used in an optimization. It is not uncommon for errors in measurement of head loss to be on the same order of magnitude or larger that the actual head loss (Walski, 2000). Such values should not be used in calibration because the calibration algorithm will dutifully try to match the field observations even if they are erroneous. To ensure that head loss is adequate to exceed measurement error, it is helpful to collect data when velocities in pipes are appreciable. In some systems sized for fire protection, demands (and velocities and head losses) are so low most of the time that head loss measurements are meaningless, other than to check pressure gage elevations. Another problem that occurs when calibrating a model is that some of the parameters determined are fixed and knowable at the time the data were taken (roughness, valve status), while others are merely a random observation from a stochastic process (water use). If a C-factor is determined as 90, then that value will be true in the not to distant future. If water use during a pressure observation is determined to be 100 gpm (6.3 l/s), is that value the demand that should be used in modeling, given that it is only one observation from a distribution? The actual water determined from calibration may not be the best value to use for representing the current year status of the system. You need to decide if the water use observed during calibration is the water use that should be used as a basis for future modeling.