1 | // Copyright (C) 2005, International Business Machines |
---|
2 | // Corporation and others. All Rights Reserved. |
---|
3 | #if defined(_MSC_VER) |
---|
4 | // Turn off compiler warning about long names |
---|
5 | # pragma warning(disable:4786) |
---|
6 | #endif |
---|
7 | |
---|
8 | #include <cassert> |
---|
9 | #include <iomanip> |
---|
10 | |
---|
11 | |
---|
12 | // For Branch and bound |
---|
13 | #include "OsiSolverInterface.hpp" |
---|
14 | #include "CbcModel.hpp" |
---|
15 | #include "CoinModel.hpp" |
---|
16 | // For Linked Ordered Sets |
---|
17 | #include "CbcBranchLink.hpp" |
---|
18 | #include "OsiClpSolverInterface.hpp" |
---|
19 | |
---|
20 | #include "CoinTime.hpp" |
---|
21 | |
---|
22 | |
---|
23 | /************************************************************************ |
---|
24 | |
---|
25 | This shows how we can define a new branching method to solve problems with |
---|
26 | nonlinearities and discontinuities. |
---|
27 | |
---|
28 | We are going to solve the problem |
---|
29 | |
---|
30 | minimize abs ( 1.0/6.931 - x1*x4/x2*x3) |
---|
31 | |
---|
32 | where the variables are integral between 12 and 60. |
---|
33 | See E.Sangren, "Nonlinear Integer and Discrete Programming in |
---|
34 | Mechanical Design Optimization". Trans. ASME, J. Mech Design 112, 223-229, 1990 |
---|
35 | |
---|
36 | One could try to use logarithms to make the problem separable but that leads to a |
---|
37 | weak formulation. Instaed we are going to use linked |
---|
38 | special ordered sets. The generalization with column generation can be even more powerful |
---|
39 | but is not yet in CBC. |
---|
40 | |
---|
41 | The idea is simple: |
---|
42 | |
---|
43 | A linear variable is a convex combination of its lower bound and upper bound! |
---|
44 | If x must lie between 12 and 60 then we can substitute for x as x == 12.0*xl + 60.0*xu where |
---|
45 | xl + xu == 1.0. At first this looks cumbersome but if we have xl12, xl13, ... xl60 and corresponding |
---|
46 | xu and yl and yu then we can write: |
---|
47 | |
---|
48 | x == sum 12.0*xl[i] + 60.0* xu[i] where sum xl[i] + xu[i] == 1.0 |
---|
49 | and |
---|
50 | x*y == 12.0*12.0*xl12 + 12.0*60.0*xu12 + 13.0*12.0*xl13 + 13.0*60.0*x13 .... |
---|
51 | + 12.0*60*.0xl60 + 60.0*60.0*xu60 |
---|
52 | |
---|
53 | And now x*y is correct if x is integer and xl[i], xu[i] are only nonzero for one i. |
---|
54 | Note that this would have worked just as easily for y**2 or any clean function of y. |
---|
55 | |
---|
56 | So this is just like a special ordered set of type 1 but on two sets simultaneously. |
---|
57 | The idea is even more powerful if we want other functions on y as we can branch on all |
---|
58 | sets simultaneously. |
---|
59 | Also note that convexity requirements for any non-linear functions are not needed. |
---|
60 | |
---|
61 | So we need a new branching method to do that - see CbcBranchLink.?pp |
---|
62 | |
---|
63 | We are going to need a CbcBranchLink method to see whether we are satisfied etc and also to |
---|
64 | create another branching object which knows how to fix variables. We might be able to use an |
---|
65 | existing method for the latter but let us create two methods CbcLink and |
---|
66 | CbcLinkBranchingObject. |
---|
67 | |
---|
68 | For CbcLink we will need the following methods: |
---|
69 | Constructot/Destructor |
---|
70 | infeasibility - returns 0.0 if feasible otherwise some measure of infeasibility |
---|
71 | feasibleRegion - sets bounds to contain current solution |
---|
72 | createBranch - creates a CbcLinkBranchingObject |
---|
73 | |
---|
74 | For CbcLinkBranchingObject we need: |
---|
75 | Constructor/Destructor |
---|
76 | branch - does actual fixing |
---|
77 | print - optional for debug purposes. |
---|
78 | |
---|
79 | The easiest way to do this is to cut and paste from CbcBranchActual to get current |
---|
80 | SOS stuff and then modify that. |
---|
81 | |
---|
82 | ************************************************************************/ |
---|
83 | |
---|
84 | int main (int argc, const char *argv[]) |
---|
85 | { |
---|
86 | |
---|
87 | OsiClpSolverInterface solver1; |
---|
88 | |
---|
89 | /* |
---|
90 | We are going to treat x1 and x2 as integer and x3 and x4 as a set. |
---|
91 | We define two new variables y1 == x1*x4 and y2 == x2*x3. |
---|
92 | We define a variable z == x1*x4/x2*x3 so y2*z == y1 |
---|
93 | (we will treat y2 as a set) |
---|
94 | Then we have objective - minimize w1 + w2 where |
---|
95 | w1 - w2 = 1.0/6.931 - z |
---|
96 | |
---|
97 | The model would be a lot smaller if we had column generation. |
---|
98 | */ |
---|
99 | // Create model |
---|
100 | CoinModel build; |
---|
101 | // Keep values of all variables for reporting purposes even if not necessary |
---|
102 | /* |
---|
103 | z is first, then x then y1,y2 then w1,w2 |
---|
104 | then y1 stuff, y2 stuff and finally y2 -> z stuff. |
---|
105 | For rows same but 2 per y then rest of z stuff |
---|
106 | */ |
---|
107 | int loInt=12; |
---|
108 | int hiInt=60; |
---|
109 | int ybaseA=5, ybaseB=9, ylen=hiInt-loInt+1; |
---|
110 | int base = ybaseB+2*2*ylen; |
---|
111 | int yylen = hiInt*hiInt-loInt*loInt+1; |
---|
112 | int zbase = 10; |
---|
113 | int i; |
---|
114 | // Do single variables |
---|
115 | double value[] ={1.0,1.0}; |
---|
116 | int row[2]; |
---|
117 | /* z - obviously we can't choose bounds too tight but we need bounds |
---|
118 | so choose 20% off as obviously feasible. |
---|
119 | fastest way to solve would be too run for a few seconds to get |
---|
120 | tighter bounds then re-formulate and solve. */ |
---|
121 | double loose=0.2; |
---|
122 | double loZ = (1-loose)*(1.0/6.931), hiZ = (1+loose)*(1.0/6.931); |
---|
123 | row[0]=0; // for reporting |
---|
124 | row[1]=zbase+1; // for real use |
---|
125 | build.addColumn(2,row,value,loZ, hiZ, 0.0); |
---|
126 | // x |
---|
127 | for (i=0;i<4;i++) { |
---|
128 | row[0]=i+1; |
---|
129 | build.addColumn(1,row,value,loInt, hiInt,0.0); |
---|
130 | // we don't need to say x2, x3 integer but won't hurt |
---|
131 | build.setInteger(i+1); |
---|
132 | } |
---|
133 | // y |
---|
134 | for (i=0;i<2;i++) { |
---|
135 | // y from x*x, and convexity |
---|
136 | row[0]=ybaseA+2*i; |
---|
137 | if (i==0) |
---|
138 | row[1]=zbase+2; // yb*z == ya |
---|
139 | else |
---|
140 | row[1]=zbase-1; // to feed into z |
---|
141 | build.addColumn(2,row,value,loInt*loInt, hiInt*hiInt,0.0); |
---|
142 | // we don't need to say integer but won't hurt |
---|
143 | build.setInteger(ybaseA+i); |
---|
144 | } |
---|
145 | // skip z convexity put w in final equation |
---|
146 | row[0]=zbase+1; |
---|
147 | build.addColumn(1,row,value,0.0,1.0,1.0); |
---|
148 | value[0]=-1.0; |
---|
149 | build.addColumn(1,row,value,0.0,1.0,1.0); |
---|
150 | // Do columns so we know where each is |
---|
151 | for (i=ybaseB;i<base+(2*yylen);i++) |
---|
152 | build.setColumnBounds(i,0.0,1.0); |
---|
153 | // Now do rows |
---|
154 | // z definition |
---|
155 | build.setRowBounds(0,0.0,0.0); |
---|
156 | for (i=0;i<yylen;i++) { |
---|
157 | // l |
---|
158 | build.setElement(0,base+2*i,-loZ); |
---|
159 | // u |
---|
160 | build.setElement(0,base+2*i+1,-hiZ); |
---|
161 | } |
---|
162 | // x |
---|
163 | for (i=0;i<2;i++) { |
---|
164 | int iVarRow = 1+i; |
---|
165 | int iSetRow = 4-i; // as it is x1*x4 and x2*x3 |
---|
166 | build.setRowBounds(iVarRow,0.0,0.0); |
---|
167 | build.setRowBounds(iSetRow,0.0,0.0); |
---|
168 | int j; |
---|
169 | int base2 = ybaseB + 2*ylen*i; |
---|
170 | for (j=0;j<ylen;j++) { |
---|
171 | // l |
---|
172 | build.setElement(iVarRow,base2+2*j,-loInt); |
---|
173 | build.setElement(iSetRow,base2+2*j,-loInt-j); |
---|
174 | // u |
---|
175 | build.setElement(iVarRow,base2+2*j+1,-hiInt); |
---|
176 | build.setElement(iSetRow,base2+2*j+1,-loInt-j); |
---|
177 | } |
---|
178 | } |
---|
179 | // y |
---|
180 | for (i=0;i<2;i++) { |
---|
181 | int iRow = 5+2*i; |
---|
182 | int iConvex = iRow+1; |
---|
183 | build.setRowBounds(iRow,0.0,0.0); |
---|
184 | build.setRowBounds(iConvex,1.0,1.0); |
---|
185 | int j; |
---|
186 | int base2 = ybaseB + 2*ylen*i; |
---|
187 | for (j=0;j<ylen;j++) { |
---|
188 | // l |
---|
189 | build.setElement(iRow,base2+2*j,-loInt*(j+loInt)); |
---|
190 | build.setElement(iConvex,base2+2*j,1.0); |
---|
191 | // u |
---|
192 | build.setElement(iRow,base2+2*j+1,-hiInt*(j+loInt)); |
---|
193 | build.setElement(iConvex,base2+2*j+1,1.0); |
---|
194 | } |
---|
195 | } |
---|
196 | // row that feeds into z and convexity |
---|
197 | build.setRowBounds(zbase-1,0.0,0.0); |
---|
198 | build.setRowBounds(zbase,1.0,1.0); |
---|
199 | for (i=0;i<yylen;i++) { |
---|
200 | // l |
---|
201 | build.setElement(zbase-1,base+2*i,-(i+loInt*loInt)); |
---|
202 | build.setElement(zbase,base+2*i,1.0); |
---|
203 | // u |
---|
204 | build.setElement(zbase-1,base+2*i+1,-(i+loInt*loInt)); |
---|
205 | build.setElement(zbase,base+2*i+1,1.0); |
---|
206 | } |
---|
207 | // and real equation rhs |
---|
208 | build.setRowBounds(zbase+1,1.0/6.931,1.0/6.931); |
---|
209 | // z*y |
---|
210 | build.setRowBounds(zbase+2,0.0,0.0); |
---|
211 | for (i=0;i<yylen;i++) { |
---|
212 | // l |
---|
213 | build.setElement(zbase+2,base+2*i,-(i+loInt*loInt)*loZ); |
---|
214 | // u |
---|
215 | build.setElement(zbase+2,base+2*i+1,-(i+loInt*loInt)*hiZ); |
---|
216 | } |
---|
217 | // And finally two more rows to break symmetry |
---|
218 | build.setRowBounds(zbase+3,-COIN_DBL_MAX,0.0); |
---|
219 | build.setElement(zbase+3,1,1.0); |
---|
220 | build.setElement(zbase+3,4,-1.0); |
---|
221 | build.setRowBounds(zbase+4,-COIN_DBL_MAX,0.0); |
---|
222 | build.setElement(zbase+4,2,1.0); |
---|
223 | build.setElement(zbase+4,3,-1.0); |
---|
224 | solver1.loadFromCoinModel(build); |
---|
225 | // To make CbcBranchLink simpler assume that all variables with same i are consecutive |
---|
226 | |
---|
227 | double time1 = CoinCpuTime(); |
---|
228 | solver1.initialSolve(); |
---|
229 | solver1.writeMps("bad"); |
---|
230 | CbcModel model(solver1); |
---|
231 | model.solver()->setHintParam(OsiDoReducePrint,true,OsiHintTry); |
---|
232 | model.solver()->setHintParam(OsiDoScale,false,OsiHintTry); |
---|
233 | |
---|
234 | CbcObject ** objects = new CbcObject * [3]; |
---|
235 | /* Format is number in sets, number in each link, first variable in matrix) |
---|
236 | and then a weight for each in set to say where to branch. |
---|
237 | In this case use NULL to say 0,1,2 ... |
---|
238 | Finally a set number as ID. |
---|
239 | */ |
---|
240 | objects[0]=new CbcLink(&model,ylen,2,ybaseB,NULL,0); |
---|
241 | objects[0]->setPriority(10); |
---|
242 | objects[1]=new CbcLink(&model,ylen,2,ybaseB+2*ylen,NULL,0); |
---|
243 | objects[1]->setPriority(20); |
---|
244 | objects[2]=new CbcLink(&model,yylen,2,base,NULL,0); |
---|
245 | objects[2]->setPriority(1); |
---|
246 | model.addObjects(3,objects); |
---|
247 | for (i=0;i<3;i++) |
---|
248 | delete objects[i]; |
---|
249 | delete [] objects; |
---|
250 | model.messageHandler()->setLogLevel(1); |
---|
251 | // Do complete search |
---|
252 | |
---|
253 | model.setDblParam(CbcModel::CbcMaximumSeconds,1200.0); |
---|
254 | model.setDblParam(CbcModel::CbcCutoffIncrement,1.0e-8); |
---|
255 | model.branchAndBound(); |
---|
256 | |
---|
257 | std::cout<<"took "<<CoinCpuTime()-time1<<" seconds, " |
---|
258 | <<model.getNodeCount()<<" nodes with objective " |
---|
259 | <<model.getObjValue() |
---|
260 | <<(!model.status() ? " Finished" : " Not finished") |
---|
261 | <<std::endl; |
---|
262 | |
---|
263 | |
---|
264 | if (model.getMinimizationObjValue()<1.0e50) { |
---|
265 | |
---|
266 | const double * solution = model.bestSolution(); |
---|
267 | int numberColumns = model.solver()->getNumCols(); |
---|
268 | double x1=solution[1]; |
---|
269 | double x2=solution[2]; |
---|
270 | double x3=solution[3]; |
---|
271 | double x4=solution[4]; |
---|
272 | printf("Optimal solution %g %g %g %g\n",x1,x2,x3,x4); |
---|
273 | for (int iColumn=0;iColumn<numberColumns;iColumn++) { |
---|
274 | double value=solution[iColumn]; |
---|
275 | if (fabs(value)>1.0e-7) |
---|
276 | std::cout<<iColumn<<" "<<value<<std::endl; |
---|
277 | } |
---|
278 | } |
---|
279 | return 0; |
---|
280 | } |
---|