Causal determination

Causal reasoning has been advocated as a central tenet of BBN modelling (Pearl 88). From this viewpoint the topology of a BBN is said to reflect the causal structure of the real world and as such provides an explanatory framework for inference and prediction. This interpretation of causality has its roots in Locke’s epistemological view of causality as a property of the world rather than simply a way of describing the world. Here causation is not only a component of experience but also an objective form of inter-dependence between objects and events. This view contrasts with that of Hume which has enjoyed considerable influence in modern science and holds that causation is a purely mental construct where “causes” and “effects” are merely connected rather than produced one by the other. Philosophical debates about the true nature of causality are outside the scope of this work but the interested reader might wish to turn to (Bunge 79) for a detailed discussion of the pros and cons of each school of thought. It is sufficient to say that for our purposes the idea that causation as a property of the real world holds some attraction since we are interested in predicting/explaining properties of real systems.

Causal structuring in BBNs helps embody theories about how the world operates as well as encoding uncertainties about its operation. Such a view contrasts sharply with classical statistical analysis where correlation between variables indicates mere phenomenological connection rather than causation. Any causal explanation is then regarded as purely subjective. It is this key distinction that gives BBNs the edge over statistical analysis. Structuring models using causal ideas is more natural for the expert since descriptions of reasoning are easily explained by comparison with reality. On the other hand in statistical modelling statements about the world can often be hidden in mathematical parameters which have no direct physical meaning. Likewise BBNs offer an advantage over rule based expert systems in that causality offers different experts a common medium to compare their reasoning. Rule based expert systems encode the mental rules of the expert independently of whether these rules are consistent with the real world.