From: Philippe Bonnardel <100604.1637@CompuServe.COM> Date: 29 Apr 96 19:42:04 EDT Subject: Global optimization test result Function: Result: No. function calls: Error Tolerance: PERM(4,50) 2.4e-07 ~75 000 1.0e-07 PERM(4,0.5) 1.9e-07 ~350 000 1.0e-07 PERM(4,0.005) 2.2e-08 ~750 000 1.0e-07 PERMO(4,10) 4.5e-10 ~65 000 1.0e-10 PERMO(4,1) 9.3e-12 ~200 000 1.0e-10 PERMO(4,0.1) 2.3e-11 ~200 000 1.0e-10 TRID(10) -210.0 ~4 000 1.0e-05 TRID(25) -2900.0 ~50 000 1.0e-05 TRID(50) -22050.0 ~600 000 1.0e-05 TRID(75) -73075.0 ~2 800 000 1.0e-05 TRID(100) -171600.0 ~5 500 000 1.0e-05 POWERSUM(8,18,44,114) 3.2e-08 ~3 000 000 1.0e-08 The standard bounds were used. Method : Simulated annealing (a variation of the Goffe code) ---------------- Here are the results obtained with the Amebsa algorithm. Function: Result: No. function calls: Error tolerance: PERM(4,50) 3.2e-16 1331 1.0e-15 PERM(4,0.5) 8.3e-17 1275 1.0e-15 PERM(4,0.005) 2.9e-16 956 1.0e-15 PERM(10,10^9) 1.1 51888 1.0e-05 PERMO(4,10) 3.3e-16 480 1.0e-15 PERMO(4,1) 1.1e-16 553 1.0e-15 PERMO(4,0.1) 1.8e-16 581 1.0e-15 PERMO(10,100) 1.3e-02 4300 1.0e-05 TRID(4) -16.0 144 1.0e-05 TRID(10) -210.0 1096 1.0e-05 TRID(20) -1520.0 4774 1.0e-05 POWERSUM(8,18,44,144) 5.0e-13 4716 1.0e-15 The standard bounds were only used for the initialization. Method : Simplex + Simulated annealing Algorithm : Amebsa Authors : W.H. Press, S.A. Teukolsky, W.T. Vetterling, B.P. Flannery. Book : Numerical Recipes in C : The Art of Scientific Computing - 2nd ed. ISBN : 0-521-43108-5 But for Amebsa, the best result and/or the number of function calls often depend on the random numbers sequence (sometimes, big variations !).