Investigate significant deviations in gurobi results given same data
brookslogan opened this issue · comments
When running many fits along the lines of #12, sometimes, gurobi will sometimes (looks infrequent or rare) give significantly different results for the same inputs. Again, the problems seem to arise on extreme quantiles levels; in the example I have at hand, all differences they were in predictions for the two lowest quantile levels, and were isolated to output from a single pair of fits from a selection of 6 fits. Differences in predictions were very substantial in these cases, with 60% involving one prediction being >=2x the other, and 25% (40% of the 60%) involving one prediction being thresholded at 0 and the other being positive.