Convexity analysis in detecting a steel plant hidden global optimum

Abstract

This case study demonstrates the value of classical analysis and to a lesser degree, system decomposition for finding a global optimum missed by a sequential linear programming scheme which converges to a non-global local minimum. The example is a 20 variable steelmaking problem in which the variable annual cost to be minimized is linear, as are all constraints except a non-convex one in each blast furnace. The sequential linear programming method gives a provenlocal minimum, although the non-convex nonlinearity prevents any proof of global optimality. The provenglobal minimum found here has a 4% lower cost. The local minimum costs only 0.2% per annum less than the rather flat global maximum, so the original local minimization only achieved about 5% of the economy possible. In the overall plant, the cost saving is over three million US$ (1972) annually.