The nodes representing hand types were originally given values which divided hands into 17 categories of final hands reported in section 3.1.2. Since any pair-low (pair of 9 or lower) hand, for example, is treated as equal to any other, BPP would bet inappropriately strongly on middling paired hands and inappropriately weakly on hands such as a pair of nines. As more than a quarter (0.26) of all hands dealt were classified as a ``Pair Low'', in many games very little useful information was being conveyed. In principle we could provide a different hand type for each poker hand that has a distinct value, since there are only finitely many of them. That finite number, however, is fairly large from the point of view of Bayesian net propagation; for example, there are already 156 differently valued Full Houses. It was opted for a modest additional refinement, moving from 17 hand types to 24 types, subdividing the most frequently occurring pair-low and pair-med to classify each paired hand separately.
This refinement appears to be sufficient to achieve significant performance gains. Further refinement may be beneficial in future when hand classification again becomes a bottleneck in performance, but until such time, it appears little improvement can be gained by refining hands further.