23 March 2002 CPAIOR ‘02 1
A New Generationof
Mixed-Integer Programming Codes
Mary Fenelon, Zongao Gu, Javier Lafuente, Ed Rothberg, Roland WunderlingILOG, Inc
Robert E. Bixbyand
23 March 2002 CPAIOR ‘02 2
Outline
• LP– Overview– Computational results
• MIP– Examples– Historical view– Features– Computational results– One more example -- future
23 March 2002 CPAIOR ‘02 3
LP
23 March 2002 CPAIOR ‘02 4
A linear program (LP) is an optimization problem of the form
LP
uxl
bAxToSubject
xcMaximize T
23 March 2002 CPAIOR ‘02 5
What’s the biggest change?
• 1988 – One algorithm for LP– Primal simplex (Dantzig, 1947)
• Today – Three algorithms for LP– Primal simplex– Dual simplex (Lemke, 1954)– Barrier (Karmarkar, 1984)
LP
23 March 2002 CPAIOR ‘02 6
Progress: 1988 – Present
• Algorithms– Simplex algorithms 960x– Simplex + barrier algorithms 2360x
• Machines– Simplex algorithms 800x– Barrier algorithms 13000x
LP
Total: Over 2000000x
23 March 2002 CPAIOR ‘02 7
Algorithm Comparison
Size Prim/ Dual/ Bar/ (#rows) #Models Dual Bar Simp
> 0 680 1.5 1.1 1.1 > 10000 248 2.0 1.0 1.2 >100000 73 2.1 1.6 0.9
LP
Key: Ratio > 1 means denominator better
23 March 2002 CPAIOR ‘02 8
MIP
23 March 2002 CPAIOR ‘02 9
A mixed-integer program (MIP) is an optimization problem of the form
LP
integerallorsome j
T
x
uxl
bAxtoSubject
xcMaximize
23 March 2002 CPAIOR ‘02 10
Example 1: LP still can be HARDSGM: Schedule Generation
Model157323 rows, 182812 columns, 6348437 nzs
• LP relaxation at root node:– Barrier: Solve time estimate 3-6 days.– Primal steepest edge: 64,000 seconds
• Branch-and-bound– 368 nodes enumerated, infeasibility reduced by 3x.– Time: 2 weeks.
• Currently “solved” by decomposition.
MIP
23 March 2002 CPAIOR ‘02 11
Example 2: MIP really is HARD
A Customer Model: 44 cons, 51 vars, 167 nzs, maximization 51 general integer variables (inf. bounds)
Branch-and-Cut: Initial integer solution -2186.0 Initial upper bound -1379.4
…after 120,000 seconds, 32,000,000 B&C nodes, 5.5 Gig tree Integer solution and bound: UNCHANGED
MIP
23 March 2002 CPAIOR ‘02 12
Example 2 (cont.): Avoid structures like
Maximize x + y + zSubject To 2 x + 2 y 1 z = 0 x free y free x,y integer
MIP
Note: This problem can be solved in several ways• Euclidean reduction on the constraint [Presolve]• Removing z=0, objective is integral [Presolve]• Bounds on variables (==> local cuts)
However: Branch-and-bound cannot solve!
23 March 2002 CPAIOR ‘02 13
• Model description: – Weekly model (repeated), daily buckets: Objective to
minimize end-of-day inventory.– Production (single facility), inventory, shipping (trucks),
wholesalers (demand known)
• Initial modeling phase– Simplified prototype + complicating constraints
(consecutive day production, min truck constraints)– RESULT: Couldn’t get good feasible solutions.
• Decomposition approach– Talk to manual schedulers: They first decide on
“producibles” schedule. Simulate using Constraint Programming.
– Fixed model: Fix variables and run MIP
Example 3: A typical situation – Supply-chain scheduling
MIP
23 March 2002 CPAIOR ‘02 14
Integer optimal solution (0.0001/0): Objective = 1.5091900536e+05Current MIP best bound = 1.5090391809e+05 (gap = 15.0873)Solution time = 3465.73 sec. Iterations = 7885711 Nodes = 489870 (2268)
CPLEX 5.0:
Implied bound cuts applied: 55Flow cuts applied: 200
Integer optimal solution (0.0001/1e-06): Objective = 1.5091904146e+05Current MIP best bound = 1.5090843265e+05 (gap = 10.6088, 0.01%)Solution time = 1.53 sec. Iterations = 3187 Nodes = 58 (2)
CPLEX 6.5:
Supply-chain scheduling (continued): Solving the fixed model
MIP
Original model: Now solves in 2 hours(20% improvement in solution quality)
23 March 2002 CPAIOR ‘02 15
Computational History:1950 –1998
• 1954 Dantzig, Fulkerson, S. Johnson: 42 city TSP– Solved to optimality using
cutting planes and solving LPs by hand
• 1957 Gomory– Cutting plane algorithm: A
complete solution• 1960 Land, Doig, 1965
Dakin– B&B
• 1971 MPSX/370, Benichou et al.
• 1972 UMPIRE, Forrest, Hirst, Tomlin (Beale)– SOS, pseudo-costs, best
projection, …
• 1972 – 1998 Good B&B remained the state-of-the-art in commercial codes, in spite of– 1973 Padberg– 1974 Balas (disjunctive
programming)– 1983 Crowder, Johnson,
Padberg: PIPX, pure 0/1 MIP– 1987 Van Roy and Wolsey:
MPSARX, mixed 0/1 MIP– Grötschel, Padberg, Rinaldi
…TSP (120, 666, 2392 city models solved)
MIP
23 March 2002 CPAIOR ‘02 16
1998…A new generation of MIP codes
• Linear programming– Stable, robust performance
• Variable/node selection– Probing on dives (strong
branching)
• Primal heuristics – 8 different tried at root (one
new one is local improvement)
– Retried based upon success
• Node presolve– Fast, incremental bound
strengthening
• Presolve– Probing in constraints: xj ( uj) y, y = 0/1 xj ujy (for all j)
• Cutting planes– Gomory, knapsack covers,
flow covers, mix-integer rounding, cliques, GUB covers, implied bounds, path cuts, disjunctive cuts
– New features• Extensions of
knapsacks• Aggregation for flow
covers and MIR
MIP
23 March 2002 CPAIOR ‘02 17
Gomory Mixed Cut• Given y, xj Z+, and
y + aijxj = d = d + f, f > 0
• Rounding: Where aij = aij + fj, define
t = y + (aijxj: fj f) + (aijxj: fj > f) Z• Then
(fj xj: fj f) + (fj-1)xj: fj > f) = d - t• Disjunction:
t d (fjxj : fj f) f
t d ((1-fj)xj: fj > f) 1-f• Combining:
((fj/f)xj: fj f) + ([(1-fj)/(1-f)]xj: fj > f) 1
MIP
23 March 2002 CPAIOR ‘02 18
Computing Gomory Mixed Cuts
1. Make a an ordered list of “sufficiently” fractional variables.
2. Take the first 100. Compute corresponding tableau rows. Reject if coeff. range too big.
3. Add to LP.4. Repeat twice.5. Computed only at root. Slack cuts
purged at end of root computation.
MIP
23 March 2002 CPAIOR ‘02 19
Computational Results I: 964 models
• Ran for 100,000 seconds (defaults)– CPLEX 5.0: Failed to solve 426 (44%)– CPLEX 8.0: Failed to solve 254 (26 %)
• Among not solved (with CPLEX 8.0)– 109 had gap < 10%– 65 had no integral solution (7%)
• With “mip emphasis feasibility”: 19 found no feasible solution (2.0%)
MIP
23 March 2002 CPAIOR ‘02 20
Computational Results II: 651 models
(all solvable)
• Ran for 100,000 seconds (defaults)• Relative speedups:
– All models (651): 12x– CPLEX 5.0 > 1 second (447): 41x– CPLEX 5.0 > 10 seconds (362):
87x– CPLEX 5.0 > 100 seconds (281):
171x
MIP
23 March 2002 CPAIOR ‘02 21
Computational Results III: 78 Models
CPLEX 5.0 not solvableCPLEX new solvable < 1000 seconds
• No cuts 33.3x• No presolve 7.7x• Old variable selection 2.7x• CPLEX 5.0 presolve 2.6x• Node presolve 1.3x• Heuristics 1.1x
• Dive probing 1.1x
MIP
23 March 2002 CPAIOR ‘02 22
Example: Network Design (France Telecom – C. Le Pape & L. Perron)
• Construct a virtual private networks– Determine routes– Determine capacities
• 6 additional constraints: 64 = 26
possibilities1. Limit traffic at each node2. Limit # of arcs in and out of nodes3. Limit # of jumps4. Symmetry constraint5. 2-line constraint6. Security constraint
• 10 minute solve time limit
MIP
23 March 2002 CPAIOR ‘02 23
CPLEX solve times (France Telecom):
GUB cover cuts applied: 328Cover cuts applied: 1290Gomory fractional cuts applied: 2
Integer optimal solution: Objective = 1.6461200000e+05Solution time = 525181.51 sec. Iterations = 469805329 Nodes = 3403990
10 Minutes: 34% gap
CPLEX 8.0: Default
GUB cover cuts applied: 803Cover cuts applied: 807Gomory fractional cuts applied: 12
Integer optimal, tolerance (0.0001/1e-06) : Objective = 1.6461200000e+05Current MIP best bound = 1.6459555512e+05 (gap = 16.4449, 0.01%)Solution time = 9275.43 sec. Iterations = 26528289 Nodes = 241051 (4219)
10 Minutes: 10% gap
CPLEX 8.0: Tuned with “mip emphasis” (4 processors)
MIP
23 March 2002 CPAIOR ‘02 24
Faster integral solutions (France Telecom) :
• Constraint Programming Approach– Build greedy initial solution.– “Sliced based search” to improve solution (Goals &
propogation)
• Results compared to CP approach– 33 cases CPLEX gives no integral solution– 31 remaining: 18 in which CPLEX produces better solutions
• Now possible in CPLEX– Advanced presolve (to use original problem representation)– Concert technology (ILOG Solver-style modeling)– Implemented local cuts– Implemented ILOG Solver-style goals
MIP