Conclusion and Future Work

The use of a translation algorithm that converts the compositional modelling problem into an aDCSP allows criteria to be formalised. More importantly, it also enables efficient, existing and future, aDCSP solution techniques to be effectively applied to solving compositional modelling problems.

The extension of the aDCSPs with (order-of-magnitude) preferences (to
form aDPCSPs) also permits the incorporation of softer requirements in
the compositional modelling problem. In this paper,
order-of-magnitude preferences have been employed to express the
appropriateness of alternative model types for certain phenomena.
While such considerations may be described by hard constraints in the
physical systems domain^{3}, they are more
subjective in less understood problem domains, such as the ecological
modelling domain. The approach presented herein provides a means to
capture and represent the subtlety of the flexible model design
decisions.

The theoretical ideas presented in this article have been applied to real-world ecological modelling problems. In this paper, it has been demonstrated how the resultant compositional modeller can be employed to create a repository of population dynamics models. The approach has also been applied to automated model construction of large and complex ecosystems such as the MODMED model of Mediterranean vegetation [23], as reported by Keppens [18].

There are some practical and theoretical issues that need to be addressed, however. On the practical side, the types of ecological model design decisions, as represented by the assumptions and assumption classes, and as supported by the inference mechanisms, should be extended. Ecological systems tend to involve interrelated populations of individuals, instead of functional compositions of individual components as with physical systems. One particularly important type of design decision in ecological modelling is therefore granularity. This requires the introduction of novel representation formalisms and inference mechanisms such as aggregation and disaggregation. Initial work for considering populations as single entities and for dividing such entities into sub-populations when necessary has been carried out [19]. Integration of such work into the present aDPCSP framework requires further investigation.

On the theoretical side, the analysis of the complexity of the present approach is rather informal. Much remains to be done in this regard, especially when comparing to the complexity of existing compositional modellers. For this comparison, additional work will be required to adapt the current translation procedure to suit existing compositional modelling problems. Most compositional modellers are of exponential complexity, however. As they employ problem-specific solution algorithms, little is known about opportunities for improving their efficiency. This work hopes to be a first step toward further understanding this important issue.