System Dynamics with Bayesian Belief Networks

["Andreas Gregoriades"]

Hi everyone, I would like to ask if anybody has done any work in integrating system dynamics with Bayesian belief networks (BBN). When I say integrate I mean having BBN models influencing variables in the SD model. Because BBN are more appropriate in situations where there is uncertainty and risk, I believe would be a good idea to integrate them with system dynamics modelling. What would you think would be the most important advantage of such an integration. Regards Antoni From: "Andreas Gregoriades" <andreas@co.umist.ac.uk>

[Carlos Ariza]

At PA Consulting we have developed several models using Bayesian Networks in areas such as fraud detection and drug discovery. We have also been in conversations with clients to implement "hybrid" systems that utilize results from both BBNs and SD models. The current thinking on market modeling is to use an SD model to calculate "macro" market variables, and use BBNs to represent customer choice at the individual product level (which is a process surrounded by substantial uncertainty). True integration would be harder; it would require reading results from a BBN at every DT in the simulation, and using those results to modify variables in the SD model. While this is possible with some API programming in most platforms, I guess that for an SD model that has more than 50 variables or so the simulation time may become prohibitive (of course, Moores law is still operating...) In the literature for BBNs youll find descriptions of "Dynamic Bayesian Networks" which are BBNs replicated over time and connected. With some development, this approach could replace simpler SD models where many cause-effect relationships between variables are probabilistic and uncertain. But for larger models it would still be impractical. One of the limitations of current BBN implementations is that continuous variables need to be "discretized", which results in some loss of accuracy. In a model with many time steps, this error would accumulate and materially distort the results. So rather than integrating these models at the simulation step level, I lean towards using the methodologies in complementary ways. Longer- term dynamic problems that can be analyzed at a higher level of aggregation should be approach using System Dynamics. Uncertainty about cause-effect relationships is best dealt with using sensitivity analysis. BBNs work much better for decision making under uncertainty at a more granular level. Carlos Ariza From: Carlos Ariza <Carlos.Ariza@paconsulting.com>