The performance of inference depends upon a number of factors. The most dominating factors are (assuming that all nodes are discrete):
- the number of nodes in the model,
- the number of states of the nodes,
- the number of parents of each node,
- the density of the graph,
- the lengths of any undirected cycles in the graph.
These are properties of the model. In addition it is necessary to take other factors into account such as the structure of the evidence, the number of queries performed on the model, the algorithm used for inference, the triangulation of the graph, e.t.c.
Furthermore, when the efficiency of inference is not sufficient it often possible to change the structure of the model, to use different algorithms for triangulation, using memory backups of clique tables, e.t.c. to improve performance of inference.
Thus, in principle, it is impossible to make a general statement about the performance of an inference engine without knowing all details about the model. Another question is the definition of efficiency. If inference takes one second, this may be fast when only a single query is to be performed, but it is very slow when millions of queries are to be performed.
