1. 1 Simulations: Basic Procedure 1. Generate a true probability distribution, and a set of base conditionals (premises), with associated lower probability bounds. 2. Determine which derived conditionals (conclusions) each system is able to derive from the given base conditionals (including lower probability bounds for these conclusions). 3. Assign scores to each system, by comparing the lower probability bounds for the derived conditionals with the true probabilities.

2. 2 Language Restriction We adopted a simple language with four binary variables: a, b, c, and d We only considered conditionals with conjuctive antecedents and consequents (and no repetition of atoms), yielding 464 conditionals: 64 of the form  x  y  z  w 96 of the form  x  y  z 48 of the form  x  y 96 of the form  x  y  z  w 96 of the form  x  y  z 64 of the form  x  y  z  w

3. 3 The probability distributions… are fixed by randomly setting the following sixteen values:  (a),  (b|a),  (b|  a),  (c|a  b),  (c|a  b),  (c|  a  b),  (c|  a  b),  (d|a  b  c),  (d|a  b  c),  (d|a  b  c),  (d|a  b  c),  (d|a  b  c),  (d|  a  b  c),  (d|  a  b  c),  (d|  a  b  c), and  (d|  a  b  c).

4. 4 Errors We say that a system has made an error when the lower probability bound for a derived conditional exceeds the true probability.

5. 5 Scoring I The advantage-compared-to-guessing score for derived conditionals: Score AvG ( C  r D,  ) = 1/3  |r   (D|C)|.

6. 6 Scoring II The price-is-right score for derived conditionals: Score PIR ( C  r D,  ) = r, if r   (D|C), = 0, otherwise. = 0, otherwise.

7. 7 Scoring III The subtle-price-is-right score for derived conditionals: Score sPIR ( C  r D,  ) = r, if r   (D|C), =  (D|C)  r, otherwise. =  (D|C)  r, otherwise.

8. 8 Minimum Probability for Base Conditionals For each simulation, we required that the probability for each base conditional (premise) exceeded some fixed value, s.

9. 9 Simulations with Randomly Selected Base Conditionals ►  was determined at random, for each run. ► We varied the number, n, of base conditionals. (n = 2, 3, 4, 5, or 6) ► We varied, s, the minimal probability of base conditionals. (s = 0.5, 0.6, 0.7, 0.8, 0.9, or 0.9999) ► The base conditionals, for each run, were selected at random, from among those conditionals whose associated probability was at least s. ► We ran each combination of s and n one thousand times.

10. 10 Simulations with Predetermined Base Conditionals ►  was determined at random, for each run. ► For each set of predetermined base conditionals, we ran each combination of s one thousand times. (s = 0.5, 0.6, 0.7, 0.8, 0.9, and 0.9999)

11. 11 The Data

12. 12 Observations (Table 1) 1. Systems O is very conservative, and licences very few inferences. [YELLOW]

13. 13 Observations (Table 2) 3. System Z almost always outperformed the other systems by the AvG measure (even though the AvG measure sometimes punishes non-errors). [YELLOW] 4. System QC outscored all of the systems by the PIR measure (save where s = 0.9999), though system Z also scores well. [BLUE]

14. 14 Observations (Table 2) 5. System Z generally achieved positive sPIR scores, and outscored all of the systems by this measure (save in the case where s = 0.9999). [GREEN] 6. QC usually obtained negative sPIR scores, due to frequent errors.

15. 15 Observations (Table 3) 7. Where s is fixed, increasing the number of base conditionals tends to increase the scores (for all measures) for systems O, P, and Z (but not for QC [YELLOW]).

16. 16 Observations (Table 4) 8. For systems O, P, and Z, the score earned per inference tends to increase for higher values of s (exceptions in [YELLOW]).

17. 17 Conclusions 1. System P offers the same security as system O, and licences more inferences. 2. System Z licences far more inferences than system P, and generally infers lower probability bounds that are close to the true probability values (based on AvG scoring).

18. 18 Conclusions 3. QC inferences are too risky, and very few of the inferences sanctioned by QC, and not by Z, should be made. (Consider the QC  Z inferences, esp. sPIR scores.) 4. Of the four systems, system Z offers the best balance of safety versus inferential power.