The point of any counting strategy is to estimate the relative probabilities of the ten distinct card values. If you employed a multi-parameter counting strategy so that you had complete knowledge of all the relative probabilities, then you would be able make the best possible decisions for how to bet and how to play. If the available information is just from two separate counts, then the best least-squares estimate for the relative probabilities is the result of the contribution from the first count plus the independent (orthogonal) contribution from the second count. It’s difficult to see how you can improve on that by comparing the nonlinear outputs of overlapping input vectors.
Linear models work very well in blackjack when the right linearization is used. I think what you are doing is interesting, but unlikely to work better than well-chosen linear models for two independent counts. For example, two count index vectors C1 and C2 can be combined with weighting coefficients a1 and a2 to give a combined index vector:
C = a1*C1 + a2*C2.
There is long history of using two counts with different weights for different playing decisions (e.g., C1=hiopt1, C2=excess-7 count).
It’s likewise possible to use different weights for the betting decision in different value ranges of the counts, e.g., (high C1, high C2), (high C1, low C2), etc. Take a look at the figure on p. 213 of Griffin’s TOB. The linear estimate gives a rough approximation to the actual advantage, but the figure suggests that a piece-wise linear estimate would give a better fit.
Bookmarks