"... In PAGE 6: ... It is clearly observed that three special values are chosen by a major fraction of jobs. These special values are 0KB, 324KB and 2600-3000KB (slightly different values in this range depending on the clusters), and their corresponding job percentages are listed in Table6 . We can see that a large fraction (30% in average) of jobs have very small memory usage3.... ..."

"... In PAGE 10: ... The analysis of all these algorithms is relatively complicated. The current state of our knowledge is summarized in Table1 . For comparison we include also the competitive ratio of List Scheduling.... In PAGE 14: ... It essentially tries to preserve the invariant above, with some special considerations for large jobs. Thus, in this model both deterministic and randomized cases are completely solved, giving the same bounds as the randomized lower bounds in Table1 . More- over, we know that randomization does not help.... In PAGE 17: ... Joel Wein observed that for preemptive open shop scheduling there exists a 2-competitive algorithm for arbitrary m. Gerhard Woeginger and the author ob- served that the randomized lower bound from the basic model which approaches e=(e ? 1) 1:5819 (see Table1 ) can be modi ed to work for open shop, too. 4.... ..."

"... In PAGE 19: ... What about lagged effects? One source of evidence can be found in qualitative questions added to the third round of the survey, on whether current or past Trabajar participants felt that the program had improved their earning opportunities outside the program. Table10 summarizes the results. Amongst continuing participants, about half felt that the program improved their chances of getting a job; two - thirds felt that it gave them a 14 The special Trabajar module has a question on whether individuals participated in another temporary pro gram (other than Trabajar); 5% of the Trabajar sample report doing so in the third survey, as opposed to 2% in the second survey.... ..."

"... In PAGE 7: ... Hence the optimal value is just that large processing time (if there are more than two cranes), as we can complete all the other jobs while the job with large processing time is being handled. Table2 shows the test results. Our B amp;B obtains all the solutions instantly taking the advantage of our bounding functions, while CPLEX still consumes much time.... ..."

"... In PAGE 19: ... What about lagged effects? One source of evidence can be found in qualitative questions added to the third round of the survey, on whether current or past Trabajar participants felt that the program had improved their earning opportunities outside the program. Table10 summarizes the results. A high proportion of respondents felt that the program improved their chances of getting a job; roughly half felt that it gave them a marketable skill; about one quarter felt that it expanded their contacts.... ..."

"... In PAGE 3: ... In Example 1, the inference fTrader; UKg ! Discharged violates the template hfJob; Countryg ! Discharged; 75%i. To eliminate this inference, we can suppress Trader and Clerk to a special value ?Job, and suppress UK and Canada to a special value ?Country, see Table2 . Now, the new inference f?Job; ?Countryg ! Discharged has con dence 50%, less than the speci ed 75%.... ..."

"... In PAGE 2: ... In Example 1, the inference fTrader; UKg ! Discharged violates the template hfJob; Countryg ! Discharged; 75%i. To eliminate this inference, we can suppress Trader and Clerk to a special value ?Job, and suppress UK and Canada to a special value ?Country, see Table2 . Now, the new in- ference f?Job; ?Countryg ! Discharged has con dence 50%, less than the speci ed 75%.... ..."