Bayesian belief networks-communicating model predictions to non-expert end users
David Patrick Callaghan, Tom E. Baldock, Behnam Shabani, Peter J. Mumby · 2017
Utilising swell and wind wave modelling undertaken previously and extensively reported in journals, conferences and reports requires a considerable time commitment, a high level of expertise and extensive climate and reef data that are not always available when undertaking planning for management of coasts and coral reef ecosystems. Consequently, the authors seek other methods that draw out key outcomes while serving practical end user requirements and capability. Bayesian Belief Networks (BBN) have at least three attributes that make them an excellent choice to communicate complex model results to practitioners wishing to access the results of extensive wave modelling simulations. First, BBNs subsume thousands of model simulations to provide probabilistic outcomes that are easily accessible to users through a simple graphical user interface. This allows the user to access the results of potentially hundreds of hours of high-performance computing almost instantaneously. Second, by using prior probabilities for system state, a practitioner can still obtain predictions of model outcomes even when their knowledge of input parameters is incomplete. For example, a user might lack data on the depth of the lagoon or the state of coral populations but the BBN can still generate results by assuming system-wide prior probabilities for the state of each. Such priors might be informative, if based on random samples (e.g., on average, the probability of a reef habitat being complex is 0.4), or uninformative and set to equal likelihood. Third, BBNs can be run in reverse and used to identify the input conditions that are most likely to deliver a chosen outcome. This might help users ask highly directed questions of the model, such as Under which circumstances is it most likely to maintain low wave heights given a moderately fast rate of sea level rise?.