5.5.2 Result Management

To make split checking working for practical simulation it is necessary to perform some house keeping of model evaluations. This means that some of the existing evaluations -- which actually are branches of the simulation-flow-models's split tree -- need to be thrown away to avoid shortages of system resources, namely disk space and the computer's main memory. Otherwise, model evaluations would pile up and cause system failures. Therefore, we need a strategy which identifies tool evaluations that can be deleted, while at the same time enough branches are preserved to perform the efficient reuse of simulation results.

SIESTA allows a maximum number of model evaluations to exist in a split tree. As soon as a model evaluation finishes and this maximum number is exceeded, one of the existing evaluations is deleted.

Figure 5.7 illustrates this procedure. The deletion of a model evaluation is equivalent to the removal of all tool evaluations, which are associated with the evaluation, up to the first split point. The difficulty arises from the decision which of the existing model evaluations should be thrown away. SIESTA offers two strategies for this decision:

• Maximum Diversity: This strategy identifies model evaluations which share the least number of identical tool evaluations with other model evaluations. Therefore, each branch is assigned a measure which denotes its uniqueness and the one with the lowest uniqueness metric will be deleted (see Figure 5.7). As a consequence, model evaluations with rare combinations of parameters are more likely to be preserved compared to those which share more tool evaluations with others.

• Minimum Diversity: In contrast to the maximum diversity strategy, the minimum diversity strategy deletes split tree branches which have rarely been used for splitting and thus have the highest uniqueness metric. It is implicitly assumed that the corresponding model evaluations are somehow out of date and are therefore of no interest anymore.

Which of these strategies will lead to better results will strongly depend on the application in which the simulation-flow-model is used. Therefore, the user is able to select the strategy which will be in effect (maximum diversity is active by default).

For a simulation-flow-model which is used in the course of a process optimization, the minimum diversity strategy is expected to lead to better results since the optimization procedure will focus its evaluations in the vicinity of the final optimization result and will not repeat evaluations with sets of parameters which already led to discouraging results before. On the contrary, statistical analysis will be more efficient if the maximum diversity strategy is in effect.