### Table 5. Regression on Pumadyn with 32 input non-linear with high noise. Training size 64 128 256 512 1024

### Table 3. Average results for small data sets

### Table 1. Parameter estimates by approximate method I for the small data set

2000

"... In PAGE 7: ...013), v 4 5.069 See notes for Table1 . Lists of positively selected sites are the same as in Table 1 Table 3.... In PAGE 7: ...059), v 4 3.142 157 159 186 193 194 219 226 See notes for Table1 . Estimates of k are around 3.... ..."

Cited by 10

### Table 2. Parameter estimates by approximate method II for the small data set

2000

Cited by 10

### Table 4: Features selected by the GA for the small data set.

1995

"... In PAGE 10: ... The string 111000000101001111 occurred most frequently in a conver- ged population. Table4 shows all the available features, marked by a 1 if they were selected by the GA in the best string, and marked by a 0 if they were left out. This shows which features have been selected by the GA as being the most in uential for bankruptcy.... ..."

Cited by 4

### Table 1: Deviation from the best model inferred (small data set)

2003

Cited by 3

### Table 1. Performance of the FAST-MCD and FSA algorithms on some small data sets.

1999

"... In PAGE 18: ...ere all regression data sets but we ran FAST-MCD only on the explanatory variables, i.e. not using the response variable. The rst column of Table1 lists the name of each data set, followed by n and p. We stayed with the default value of h = [(n + p + 1)=2].... In PAGE 18: ...ollowed by n and p. We stayed with the default value of h = [(n + p + 1)=2]. The next column shows the number of starting (p + 1)-subsets used in FAST-MCD, which is usually 500 except for two data sets where the number of possible (p + 1)-subsets out of n was fairly small, namely ?123 = 220 and ?183 = 816, so we used all of them. The next entry in Table1 is the result of FAST-MCD, given here as the nal h-subset. By comparing these with the exact MCD algorithm of Agull o (personal communication) it turns out that these h-subsets do yield the exact global minimum of the objective function.... In PAGE 21: ... In Tables 1 and 2 we have applied the FSA algorithm to the same data sets as FAST- MCD, using the same number of starts. For the small data sets in Table1 the FSA and FAST-MCD yielded identical results. This is no longer true in Table 2, where the FSA begins to nd nonrobust solutions.... ..."

Cited by 67

### Table 1. Performance of the FAST-LTS and FSA algorithms on some small data sets.

1999

"... In PAGE 13: ... 5 Performance of FAST-LTS To get an idea of the performance of the overall algorithm, we start by applying FAST- LTS to some small regression data sets taken from (Rousseeuw and Leroy 1987). The rst column of Table1 lists the name of each data set, followed by n and p, where n is the number of observations and p stands for the number of coe cients including the intercept term. We stayed with the default value of h =[(n + p +1)=2].... In PAGE 13: ... The next column shows the number of starting p-subsets used in FAST-LTS, which is usually 500 except for two data sets where the number of possible p-subsets out of n was fairly small, namely ; 12 3 = 220 and ; 18 3 = 816, so we used all of them. The next entry in Table1 is the result of FAST-LTS, given here as the nal h-subset. By comparing these with the exact LTS algorithm of Agull o (personal communication) it turns out that these h-subsets do yield the exact global minimum of the objective function.... In PAGE 16: ... In Tables 1 and 2 wehave applied the FSA algorithm to the same data sets as FAST-LTS, using the same number of starts. For the small data sets in Table1 the FSA and FAST-LTS yielded identical results, but for the larger data sets in Table 2 the FSA obtains nonrobust solutions. This is because (1) The FSA starts from randomly drawn h-subsets H 1 .... ..."

Cited by 19

### Table 6: Prediction of best net found by the GA using the small data set

1995

Cited by 4