Knowledge representation using Bayesian Networks and Ontologies

Models are often highly complex incorporating different processes, parameters, scenarios and subjects, and thereby producing different outcome endpoints. The first recommended model development step is the construction of a conceptual model, thereby specifying and defining the processes and relationships to be covered by the model. However, model results are also often highly complex, being fractionalized or simply numerous in quantity, leading to difficulties in representation, communication and interpretation of results. We explore the use of two existing tools and their possible use in knowledge representation and visualization; Bayesian Networks using Netica and ontologies using Protégé. While visually speaking, both techniques represent knowledge or concepts through network associations between nodes, the information that underlies these representations is vastly different. Bayesian Networks capture relationships using statistical probabilities, whereas ontologies represent structured formalization of relationships. We explore the use of these two approaches in a novel problem space by representing the modelled outcomes of changed river flow regimes in the MDB to different water development and predicted climate change scenarios and the impact on meeting the watering requirements on the wetland indicator sites in the Southern Murray-Darling Basin. Evaluation of the environmental requirements of wetland indicator sites which are met under different CSIRO Sustainable Yields river flow scenarios representing 109 years of modeled river flows is carried out. As expected, the outcomes of modeling the watering requirements of the wetland sites under different river flow scenarios vary by the scenario, the site and the specific environmental requirement, where watering requirements for the wetland indicator sites are met most of the time under the ‘without development’ scenario, and only a fraction of the time (4.17%) under the baseline scenario, and less (2.08%) under dry climate scenarios. To represent the outcomes of the river flow scenarios, we present Bayesian Networks, which represent outcomes as a proportion of years where a set of environmental requirements are met, and use utility nodes to display how much additional water is required to meet site-based environmental requirements. We do this for individual wetlands and aggregate outcomes to represent asset requirements in the whole of the southern Murray Darling Basin. Likewise for the approach using ontologies, we formalize a multi-inheritance hierarchy to enable interactive representation of outcomes as defined by different criteria within the model, for example by sites, scenarios, outcomes, or watering requirements. With the ontology approach, this allows representing the outcomes from different positions within the model and observing the derived associations between individual objects based upon the relationships within the ontology model, for example the individual outcomes of a specific flow scenario on a specific wetland indicator site can be represented. Utilizing the functionality of both Bayesian Networks and ontologies in representation of model outcomes enables a deeper exploration of the underlying model data, enabling interactivity, interrogation and specific queries to be made compared to more traditional representation techniques. Ontologies provide a useful means for exploring individual relationships and associations within the data resulting through the taxonomical data structure, while Bayesian Networks enable exploring the range of specific outcomes from the different dimensions of the data. We discuss the use of Bayesian Networks and ontologies to represent this knowledge in a structured, visual and interactive manner.