Tuesday, February 5, 2008

Blind Tasting and Bayes' Theorem

Critics and wine gurus often recommend blind tasting as the best way to train one's palate. In addition, the major critics (Parker, Wine Spectator, Tanzer) taste blind to reassure us of their unwavering objectivity. However objectivity is not always the same as accuracy (for more on this, see an article by Eric Asimov), and it is not at all clear to me why blind tasting is a useful form of training. Here I will cast the act of blind tasting into a Bayesian decision model, and discuss possible implications.

First, I want to be clear on this point: I am not discussing the act of blind tasting with intent to subjectively evaluate. If your goal is to identify which wines give you most pleasure to drink, taste them as you drink them. If you drink them blind, taste them blind - otherwise, don't bother. Here I am concerned with the act of blind tasting to determine objective truths about the wine. I am taking the objectivity of tastes and tasting for granted. (For a discussion of this assumption, see Barry Smith's article in the book, Questions of Taste)

Given a set of objective statements A, and a set of observations B, we can consider the act of blind tasting as the informed decision whether to believe statements in A based on which observations in B we percieve. In sparser, mathematical terms, A and B are two disjunct sets of events. By Bayes' Theorem, the conditional probability of an event A (in the set A) given an observation B (in the set B), can be calculated by:

P(AB) = P(BA) * P(A) / P(B)

Where P(B) and P(A) are the a priori probabilities of A and B, and P(BA) is the conditional probability of B, given A. This equation articulates the truism that in order to assign a conditional probability of A based on information about B, you have to allow a non-zero probability for A in the first place. For example: to conclude that the wine has Shiraz characteristics, all the Shiraz smells in the world are insufficient if we do not first allow for the possibility that the wine contains Shiraz. The strength of our decision depends on the specificity and accuracy of our knowledge of each of the three terms on the right-hand side of the equation. Why then would would we choose to taste blind, and eliminate some of this information?

By tasting blind we aim to weaken our ability to determine P(A), and so we unconsciously "retreat" toward the default of least information: that all A's are equally possible. For example: not knowing the varietal, we would hopefully consider all varietals equally. In reality, we fail: naturally we expect Cabernet, Merlot, Chardonnay, and Riesling more that Duras, Gruner Veltliner, or Xarrello.

By reducing the distinguishing power of the set of P(A)'s, we tilt the equation to depend primarily on P(BA), which is directly derived from experience, and on P(B) - i.e. how unique our observation is. Naturally this reduces the strength of the decision, which is a mathematical reflection on our increased uncertainty when blind tasting. Unconsciously we try to compensate by seizing any scrap of information to make a particular P(A) stand out.

Mathematically, tasting blind reduces the precision of our decision-making (or to be slightly more correct, makes for a less favorable precision-recall curve). However it quite possibly increases our accuracy. If we wine tasters have a shared tendency to exaggerate unwarranted differences in P(A), blind tasting would improve our accuracy. As Emile Peynaud cautioned bluntly in his famous book, Le Gout de Vin (The Taste of Wine), "the wine taster is easily influenced."

However Bayes' equation makes the importance of P(BA) absolutely clear: tasting blind will only improve your ability to objectively analyze, if you possess accurate knowledge of the reverse probabilities. An example of this type of knowledge might be: "9 out of 10 Gewurztraminers have the taste of lychees." Note the distinction between this and P(AB): "9 out of 10 wines with the taste of lychees are made from Gewurztraminer." This type of knowledge only comes from experience. Tasting only a few Gewurztraminers is statistically insignificant, and thus for a wine novice, tasting blind is rather useless as a tool to improve accuracy. Instead, taste with prior knowledge of what the wine should taste like, and see how well your perceptions match up. As the esteemed Vayniac Joe G. wrote:

Once you can give an accurate tasting note of what a bottle should taste like from a region without tasting it, that is when blind tasting can be educational.
- WLTV forum discussion

It's like Jazz, folks. You gotta learn the notes before you improvise.

(Note: Thanks to all the Vayniacs who helped me crystallize my thoughts in the forum discussion of this topic.)