Automated modelling and scientific discovery

As mentioned previously, a compositional model repository is designed in order to compose models from a system's structure and relevant domain knowledge. As such, this approach gives rise to a potentially beneficial means to operationalise the outcomes of scientific discovery. More specifically, the resultant compositional model repositories will allow existing knowledge on model construction to be applied to unexperienced scenarios and to support investigation into situations which may be physically difficult to replicate or create but which may be synthesised in computational representations.

The present work has been applied to the vegetation component of the MODMED $n$-species model [23]. This $n$-species model offers a system dynamics representation of populations of Mediterranean vegetations and of how they are affected by populations of farm animals, climate and environmental management. The purpose of the model is to be instantiated with respect to various Mediterranean communities, and to serve as a component of a very large scale simulation that is designed to simulate the effects of various environmental policies on the Mediterranean landscape. A knowledge base containing approximately 60 model fragments and 4 property definitions has been constructed, on the basis of the most complex parts of the $n$-species model in about two man-weeks. This knowledge base can be employed to reconstruct variations of the $n$-species model to accommodate a variety of possible scenarios, as well as to examine simplifications of the original $n$-species model which exclude certain phenomena.

The compositional model repository is most closely related to the seminal work on compositional modelling [10]. That approach has a similar functionality but it is devised specifically for physical systems and relies on a component-connection formalism to represent scenarios.

Another approach which has recently been developed and applied to the ecological domain by Heller and Struss [14,15]. This work derives a system's structure from observations of its behaviour and domain knowledge. Therefore, it is able to perform diagnosis of ecological systems and therapy suggestion. Another important distinction of this work from the present study is that it presumes that each process can only be described in just one way instead of allowing multiple alternative models.

In the machine learning community, a number of approaches have been devised by Bradley, Easley and Stolle [4]; Langley et al. [22]; and Todorovski and Dzeroski [39,40] to induce sets of differential equations from a) observations of behaviour, b) domain knowledge represented in the form of hypothetical equations, and c) a description of the structure of the system. These approaches aim at scientific discovery by generalising observed behaviour into mathematical models. The specifications of the scenario and the domain knowledge in these methods are similar to those used in this article. This is especially true for the work by Langley et al. [22]; and Todorovski and Dzeroski [39,40], because that work has also been applied to population dynamics. However, the internal mechanisms of these approaches are very different as they essentially rely on exhaustive search procedures instead of constraint satisfaction techniques.