Ecosystem Analysis Using Probabilistic Relational Modeling
3. Conclusions and Future Work
Relational probabilistic modeling provides a natural framework for investigating ecological data. The large amount of observational, noisy data, often collected by multiple investigators over varying time-scales, provides a rich field for probabilistic model discovery, and relational approaches raise the level of modeling to one with which domain scientists can readily interact.
Existing synthetic variable construction methods naturally generate many variables either previously known to scientists or immediately recognized by them as scientifically relevant. At the same time, attempts to apply relational probabilistic model discovery techniques to ecological data have revealed limitations in our current synthetic variable construction methods. We are currently exploring work in data base path expressions, for example that of Van den Bussche [Van den Bussche et al., 93] and Frohn [Frohn et al., 94], as generalizations capable of expressing a more comprehensive set of synthetic variables. Key concepts include the selector and the introduction of variables (to allow subsequent reference to earlier elements in a path). We are also exploring mixed-initiative search procedures over these much larger path grammars.
1 In more recent work, supported by NSF SBIR DMI-0231961, we have developed a more comprehensive synthetic variable language grammar and automated generation capability, patent-pending.