Algorithms for dream combat strategy

I just discovered that I had already written scripts for the attack algorithms. But they don't take into account the certainty values. Instead, they simply look for the auragon color for which the attacker has the greatest probability of defeating the defender. Thus, it calculates:

Number2BNumber of:
   quotient of:
      BNumber2Number of:
         RedAuragonCount of:
            ReactingActor
      BNumber2Number of:
         PerceivedGreenAuragonCount of:
            CandidateDefender

The result of this calculation is a number representing the odds of defeating the defender. The script calculates all three combinations and selects the highest of the three. The CandidateDefender with the highest value is chosen to attack. 

Ack! I just realized that this is ENTIRELY WRONG! The important decision involves deciding which auragon the opponent is likely to deploy. We don't care what auragons we have; victory or defeat depends upon guessing their likely action. So the decision boils down to direct application of the pAuragon counts. Attack the one the opponent has most of. 

But we have to take into account the possibility that the opponent might use another. For example, suppose that Zubi has the following pValues and cValues for Camiggdo:

pRedAuragon: 2.16    Certainty: -0.5
pGreenAuragon: 1.79 Certainty +0.5
pBlueAuragon: 0.22    Certainty -0.1

Simply choosing the largest number tells me to attack with blue. But Camiggdo could well attack with green, in which case my blue would lose. So the correct figure of merit for this calculation is the probability that Camiggdo will use red MINUS the probability that Camiggdo will use green. This closely parallels a standard problem in statistics: given two values, X and Y, with respective uncertainties uX and uY, what is the probability that X exceeds Y? There's a more precise way of stating this, but I'll avoid the formality. Indeed, I see no use in applying the standard statistical formula here, as my certainty values do not reflect a normal distribution. 
Now, in the above case, the certainties do not change the fact that the BEST estimates of the pValues argue for using a blue auragon. The uncertainties do not change that fact; they serve only to make the decision less certain. Perhaps, then, the certainties should be used for deciding WHOM to attack rather than WHAT auragon to attack with. In other words, the pValues alone are used to decide the best auragon to use against each candidate, but in deciding between two likely candidates to attack, the certainties do enter into the calculation. 

No, that doesn't work, either. The first calculation yields, in effect, a probability X that color A will result in victory. If we compare two possible attacks, then we get probability X that color A will yield victory over defender 1, and probability Y that color B will yield victory over defender 2. Adding certainties to the probabilities doesn't change the final decision. 

So do certainties have zero operational value? It would appear that their only operational utility lies in incorporating them into calculations of changing one's beliefs based on inputs. That's still good enough. 

I need to think this over some more.