Dynamic Bayesian Network Inferencing for Non-Homogeneous Complex Systems

Abstract

Summary Dynamic Bayesian networks (DBNs) provide a versatile method for predictive, whole‐of‐systems modelling to support decision makers in managing natural systems subject to anthropogenic disturbances. However, DBNs typically assume a homogeneous Markov chain which we show can limit the dynamics that can be modelled especially for complex ecosystems that are susceptible to regime change (i.e. change in state transition probabilities). Such regime changes can occur as a result of exogenous inputs and/or because of past system states; the latter is known as path dependence. We develop a method for non‐homogeneous DBN inference to capture the dynamics of potentially path‐dependent ecosystems. The method enables dynamic updates of DBN parameters at each time slice in computing posterior marginal probabilities given evidence for forward inference. An approximate algorithm for forward–backward inference is also provided noting that convergence is not guaranteed in a path‐dependent system. We demonstrate the methods on a seagrass dredging case‐study and show that the incorporation of path dependence enables conditional absorption into and release from the zero state in line with ecological observations. The model helps managers to develop practical ways to manage the marked effects of dredging on high value seagrass ecosystems.

Dynamic Bayesian Networks for Evaluation of Granger Causal Relationships in Climate Reanalyses

“…Interactions between spatially separated processes, for example, may be seasonally dependent, while on longer time‐scales the possibility of climate regime changes, for example, in association with tipping points (Lenton et al., 2008), in response to anthropogenic forcing has recently become a key concern. Non‐homogeneous DBNs, in which either the graph structure, parameters or both simultaneously, are allowed to change over time, admit the possibility of modeling features such as secular trends and regime changes (Wu et al., 2018), at the cost of a significant increase in complexity in terms of model specification and inference. Here we focus on the simpler case of homogeneous models for the purposes of investigating the usefulness of Bayesian methods for assessing model uncertainty; the more complicated case of non‐homogeneous models will be described in a separate study.…”

Dynamic Bayesian Network Modeling, Learning, and Inference: A Survey

“…Content may change prior to final publication. [55] Error Rate Non-homogenous Inference [57] Error Rate Inference with constraints and sliding window [12] Time Inference with qualitative information [58] Coherence…”

Prior and Posterior Linear Pooling for Combining Expert Opinions: Uses and Impact on Bayesian Networks—The Case of the Wayfinding Model

“…Some of the practical advantages of BNs are that they are suitable for small and incomplete datasets; that they allow for structural learning; that different sources of knowledge and data types can be combined; that they can be solved analytically and hence can provide a fast, real-time response to queries; and that they can be extended to incorporate spatial and temporal dynamic processes [ 21 , 22 , 23 ].…”