/**
* run as <CODE>
* java -cp <classpath> tests.DDETest <params_file> [random_seed] [maxfuncevals]
* </CODE>. The params_file must contain lines of the following form:
*
* <ul>
* <li>class,dde.function, <fullclasspathname> mandatory, the java class name defining the
* function to be optimized, which must accept as arguments either <CODE>double[]</CODE> or
* <CODE>VectorIntf</CODE> objects.
* <li>class,dde.localoptimizer, <fullclasspathname> optional the java class name of an
* object implementing the LocalOptimizerIntf defined in the popt4jlib.GradientDescent
* package, to be used as further optimizer of the best solution found by the DE process.
* <li>dde.numdimensions, $num$ mandatory, the dimension of the domain of the function to be
* minimized.
* <li>dde.numtries, $num$ optional, the total number of "tries", default is 100.
* <li>dde.numthreads, $num$ optional, the number of threads to use, default is 1.
* <li>rndgen,$num$,$num2$ mandatory, specifies the starting random seed to use for each of the
* $num2$ threads to use (the value num2 must equal the number given in the line for
* dde.numthreads).
* <li>dde.popsize, $num$ optional, the total population size in each iteration, default is 10.
* <li>dde.w, $num$ optional, the "weight" of the DE process, a double number in [0,2], default
* is 1.0
* <li>dde.px, $num$ optional, the "crossover rate" of the DE process, a double number in [0,1],
* default is 0.9
* <li>dde.minargval, $num$ optional, a double number that is a lower bound for all variables of
* the optimization process, i.e. all variables must satisfy x_i ≥ $num$, default is
* -infinity
* <li>dde.maxargval, $num$ optional, a double number that is an upper bound for all variables
* of the optimization process, i.e. all variables must satisfy x_i ≤ $num$, default is
* +infinity
* <li>dde.minargval$i$, $num$ optional, a double number that is a lower bound for the i-th
* variable of the optimization process, i.e. variable must satisfy x_i ≥ $num$, default
* is -infinity
* <li>dde.maxargval$i$, $num$ optional, a double number that is an upper bound for the i-th
* variable of the optimization process, i.e. variable must satisfy x_i ≤ $num$, default
* is +infinity
* <li>dde.de/best/1/binstrategy, $val$ optional, a boolean value that if present and true,
* indicates that the DE/best/1/bin strategy should be used in evolving the population
* instead of the DE/rand/1/bin strategy, default is false
* <li>dde.nondeterminismok, $val$ optional, a boolean value indicating whether the method
* should return always the same value given the same parameters and same random seed(s).
* The method can be made to run much faster in a multi-core setting if this flag is set to
* true (at the expense of deterministic results) getting the CPU utilization to reach
* almost 100% as opposed to around 60% otherwise, default is false
* <li>dde.countfuncevals, $val$ optional, a boolean value indicating whether or not the process
* should be counting the number of function evaluations it performs, default is false
* <li><"dde.dmpaddress", String location> optional, if existing, specifies the location
* of a distributed Msg-Passing server that implements the basic send/recv operations as
* specified in <CODE>parallel.distributed.DActiveMsgPassingCoordinatorLongLivedConnSrv[Clt]
* </CODE> default is null
* <li><"dde.dmpport", Integer port> optional, if existing, specifies the port number of a
* distributed Msg-Passing server implementing the basic send/recv operations as specified
* in <CODE>parallel.distributed.DActiveMsgPassingCoordinatorLongLivedConnSrv[Clt]</CODE>
* default is null
* <li><"dde.dmpthisprocessid", Integer myid> optional, if existing, it indicates the id
* of this process (this is the number to use in a recvData(myid) call on the
* DActiveMsgPassingCoordinatorLongLivedConnClt object) default is null
* <li><"dde.dmpnextprocessid", Integer id> optional, if existing, it indicates the id of
* the process to which this process should be sending "migrants" to; (this is the number to
* use as the "send address" in a sendData(myid, id, data) DActiveblahblah call); default is
* null
* <li><"dde.numgensbetweenmigrations", Integer num> optional, if existing, it indicates
* the number of generations that must pass between two successive "migrations" between DDE
* island-processes; default is 10
* <li><"dde.nummigrants",Integer num> optional, if it exists, indicates how many migrants
* will be sent and received from each dde process; default is 10
* <li><"dde.reducerhost", String host> optional, if existing, it indicates the address in
* which the reducer server resides; default is null
* <li><"dde.reducerport", Integer port> optional, if existing, it indicates the address
* in which the reducer server process listens at; default is -1
* </ul>
*
* <p>Notice that in case of running DDE in a distributed manner, there are two important
* constraints:
*
* <ul>
* <li>First of all, the various processes participating in the distributed DDE process must
* have their dde.dmpthisprocessid and dde.dmpnextprocessid parameters set up so that the
* flow of migration forms an exact ring, e.g. for a 3-process DDE we have that DDE_0 sends
* migrants to DDE_1 which sends migrants to DDE_2 which sends migrants to DDE_0.
* <li>The constraint dde.numgensbetweenmigrations ≤ dde.numtries must hold.
* </ul>
*
* Otherwise, there is no way for processes to block in at least one migration cycle (and thereby
* have the DReduceSrv know the total number of processes before the final reduce operation), and
* therefore the distributed reduce operation afterwards is not guaranteed to work properly.
*
* <p>if the second optional argument is passed in, it overrides the random seed specified in the
* params_file (specified as the 1st arg in cmd-line).
*
* <p>The optional third argument, if present, overrides any max. limit set on the number of
* function evaluations allowed. After this limit, the function will always return
* Double.MAX_VALUE instead, and won't increase the evaluation count. Obviously, for this limit to
* be enforced, the "dde.countfuncevals" flag must be set to true in the params_file (1st arg in
* the command line)
*
* @param args String[]
*/
public static void main(String[] args) {
try {
long start_time = System.currentTimeMillis();
HashMap params = utils.DataMgr.readPropsFromFile(args[0]);
if (args.length > 1) {
long seed = Long.parseLong(args[1]);
RndUtil.getInstance().setSeed(seed); // updates all extra instances too!
}
if (args.length > 2) {
long num = Long.parseLong(args[2]);
params.put("maxfuncevalslimit", new Long(num));
}
FunctionIntf func = (FunctionIntf) params.get("dde.function");
if (params.containsKey("dde.countfuncevals")
&& ((Boolean) params.get("dde.countfuncevals")).booleanValue() == true) {
FunctionBase wrapper_func = new FunctionBase(func);
params.put("dde.function", wrapper_func);
func = wrapper_func;
}
DDE opter = new DDE(params);
utils.PairObjDouble p = opter.minimize(func);
VectorIntf arg = (VectorIntf) p.getArg();
System.out.print("best soln found:[");
for (int i = 0; i < arg.getNumCoords(); i++) System.out.print(arg.getCoord(i) + " ");
System.out.println("] VAL=" + p.getDouble());
// final local optimization
utils.PairObjDouble p2 = null;
LocalOptimizerIntf lasdst = (LocalOptimizerIntf) params.get("dde.localoptimizer");
if (lasdst != null) {
VectorIntf x0 = arg.newInstance(); // arg.newCopy();
params.put("gradientdescent.x0", x0);
lasdst.setParams(params);
p2 = lasdst.minimize(func);
VectorIntf xf = (VectorIntf) p2.getArg();
System.out.print("Optimized (via a GradientDescent local method) best soln found:[");
for (int i = 0; i < xf.getNumCoords(); i++) System.out.print(xf.getCoord(i) + " ");
System.out.println("] VAL=" + p2.getDouble());
}
if (func instanceof FunctionBase)
System.err.println("total function evaluations=" + ((FunctionBase) func).getEvalCount());
long dur = System.currentTimeMillis() - start_time;
double val =
(p2 == null || p2.getDouble() >= Double.MAX_VALUE) ? p.getDouble() : p2.getDouble();
System.out.println("total time (msecs): " + dur);
if (func instanceof FunctionBase)
System.out.println(
"VVV,"
+ val
+ ",TTT,"
+ dur
+ ",NNN,"
+ ((FunctionBase) func).getEvalCount()
+ ",PPP,DDE,FFF,"
+ args[0]); // for parser program to extract from output
} catch (Exception e) {
e.printStackTrace();
System.exit(-1);
}
}
/**
* run as <CODE>
* java -cp <classpath> graph.packing.DBBGASPPacker <graphfilename> <paramsfilename> [maxnumBBnodes] [numthreads]
* </CODE>. <br>
* args[0]: graph file name must adhere to the format specified in the description of the method
* <CODE>utils.DataMgr.readGraphFromFile2(String file)</CODE> <br>
* args[1]: params file name may define parameters in lines of the following form:
*
* <ul>
* <li>acchost, $string$ optional, the internet address of the DAccumulatorSrv that will be
* accumulating incumbents. Default is localhost.
* <li>accport, $num$ optional, the port to which the DAccumulatorSrv listens. Default is 7900.
* <li>cchost, $string$ optional, the internet address of the DConditionCounterSrv that will be
* listening for distributed condition-counter requests. Default is localhost.
* <li>ccport, $num$ optional, the port to which the DAccumulatorSrv listens. Default is 7899.
* <li>pdahost, $string$ optional, the internet address of the PDAsynchBatchTaskExecutorSrv that
* will be listening for distributed tasks execution requests. Default is localhost.
* <li>pdaport, $num$ optional, the port to which the asynch distributed executor server
* listens. Default is 7981.
* <li>tightenboundlevel, $num$ optional, the depth in the B&B tree constructed at which a
* stronger computation of the upper bound will be started, default is 0.
* <li>cutnodes, $boolean$ optional, if true, then when the BBQueue of nodes is full, any new
* nodes created will be discarded instead of processed on the same thread. Default is
* false.
* <li>localsearch, $boolean$ optional, if true, then when an incumbent solution is found that
* cannot be further improved, a local search kicks in to try to improve it using (unless
* there is another explicit specification) the (default) N1RXP(FirstImproving) neighborhood
* concept that basically attempts to remove a single node from the solution and then see
* how many other nodes it can add to the reduced solution. This local search can become
* quite expensive and for this reason it is only applied to final incumbent solutions in
* the B & B-tree construction process. Default is false.
* <li>class,localsearchtype, <fullclassname>[,optionalarguments] optional if present,
* will utilize in the local-search procedure the <CODE>AllChromosomeMakerClonableIntf
* </CODE> specified in the classname, constructed using the arguments specified (if
* present). For example, the line could be:
* <PRE>
* class,localsearchtype,graph.packing.IntSetN2RXPGraphAllMovesMaker,1
* </PRE>
* which would be of use with MWIS problems, for random graphs in the class C(n,p),
* producing G_{|V|,p} type random graphs. On the other hand, by default, the <CODE>
* IntSetN1RXPFirstImprovingGraphAllMovesMakerMT</CODE> moves maker applies both for 1- and
* 2-packing problems local-search, which is also better suited when solving MWIS problems
* arising from duals of disk graphs (arising from wireless ad-hoc networks etc.) Currently
* works only with MWIS (k=1) type problems and is ignored for 2-packing problems.
* <li>usemaxsubsets, $boolean$ optional, if false, then each GASP process augmenting candidate
* packings will augment these sets one node at a time, leading to the possibility that many
* active nodes in the B&B tree will represent the same packing. In such a case, a
* "recent-nodes" queue will be used to safe-guard against the possibility of having the
* same nodes created and processed within a "short" interval. Default is true.
* <li>sortmaxsubsets, $boolean$ optional, if true, then the max subsets generated in method
* <CODE>getBestNodeSets2Add()</CODE> will be sorted in descending weight order so that if
* children <CODE>BBNode*</CODE> objects are "cut", they will be the "least" heavy-weight.
* Default is false.
* <li>maxitersinGBNS2A, $num$ optional, if present and also the "usemaxsubsets" key is true,
* then the number represents the max number of iterations the <CODE>getBestNodeSets2Add()
* </CODE> method of the <CODE>DBBNode1</CODE> class will be allowed to go through. Default
* is 100000 (specified in <CODE>DBBTree</CODE> class.)
* <li>useGWMIN2criterion, $boolean$ optional, if true, then when computing the best nodes to
* consider as a partial solution is being expanded, the "GWMIN2-heuristic" criterion
* (described in Sakai et. al. 2003: "A note on greedy algorithms for the MWIS problem",
* Discr. Appl. Math., 126:313-322) will be used for nodes selection in 1-packing problems.
* Default is false.
* <li>maxnodechildren, $num$ optional, specify an upper bound on the number of children any
* node is allowed to create. Default is Integer.MAX_VALUE.
* <li>class,dbbnodecomparator, <fullclassname> optional, the full class name of a class
* implementing the <CODE>graph.packing.DBBNodeComparatorIntf</CODE> that is used to define
* the order in which B&B nodes in the tree are picked for processing. Default is <CODE>
* graph.packing.DefDBBNodeComparator</CODE>.
* <li>ff, $num$ optional, specify the "fudge factor" used in determining what constitutes the
* list of "best nodes" in the 1-packing problem (a.k.a. the MWIS problem) where it makes
* much more sense to have a "fudge factor" by which to multiply the best cost in order to
* determine if a node is "close enough" to the best cost to be included in the
* best-candidate-nodes list. Default value is <CODE>DBBNode1._ff</CODE> (currently set to
* 0.85). The smaller this value, the longer it will take for the search to complete, with
* potentially better solutions found.
* <li>minknownbound, $num$ optional, a known bound to the problem at hand, which will be used
* to fathom B&B nodes having smaller bounds than this number. Currently only applies to
* 1-packing problems. Default is -infinity.
* <li>expandlocalsearchfactor, $num$ optional, if present, then when a solution is found within
* the specified factor of the best known solution, a local search kicks in. Default is 1.0
* (only when a best solution is found does local search kicks in). Currently only applies
* to 1-packing problems.
* </ul>
*
* <br>
* args[2]: [optional] override max num nodes in params_file to create in B&B procedure
*
* <p>This implementation writes the solution in a text file called "sol.out" in the current
* directory, whose lines contain one number each, the id of each "active" node in the solution
* (id in the set {1,...graph_num_nodes}).
*
* @param args String[]
*/
public static void main(String[] args) {
try {
if (args.length < 2) {
System.err.println(
"usage: java -cp <classpath> graph.packing.DBBGASPPacker <graphfilename> <paramsfilename> [maxnumBBnodes]");
System.exit(-1);
}
long start = System.currentTimeMillis();
Graph g = DataMgr.readGraphFromFile2(args[0]);
// print out total value weight of the nodes
{
double totw = 0.0;
for (int i = 0; i < g.getNumNodes(); i++) {
Double w = g.getNode(i).getWeightValue("value");
totw += w == null ? 1.0 : w.doubleValue();
}
System.err.println("Graph total nodes' weight=" + totw);
}
HashMap params = DataMgr.readPropsFromFile(args[1]);
int maxnodes = -1;
if (args.length > 2) maxnodes = Integer.parseInt(args[2]); // override max num nodes
Graph[] graphs = null;
if (g.getNumComponents() > 1) {
// graphs = g.getGraphComponents();
System.err.println(
"Distributed BBGASPPacker does not currently support breaking graphs into disconnected components...");
System.exit(-1);
} else { // optimize when there is only one component in the graph
graphs = new Graph[1];
graphs[0] = g;
}
// now run the B&B algorithm for each sub-graph
PrintWriter pw = new PrintWriter(new FileWriter("sol.out"));
for (int j = 0; j < graphs.length; j++) {
Graph gj = graphs[j];
System.err.println(
"Solving for subgraph "
+ (j + 1)
+ " w/ sz="
+ gj.getNumNodes()
+ " (/"
+ graphs.length
+ ")");
if (gj.getNumNodes() == 3 && gj.getNumArcs() == 2) {
_totActiveNodes += 2;
++_totLeafNodes;
// figure out which node is the connecting one
int best_node_id = -1;
for (int m = 0; m < 3; m++) {
Node nm = gj.getNode(m);
if (nm.getNbors().size() == 1) {
Double nmwD = nm.getWeightValue("value");
double w = nmwD == null ? 1.0 : nmwD.doubleValue();
_totActiveNodeWeights += w;
Integer mI = (Integer) gj.getNodeLabel(m);
best_node_id = mI == null ? m : mI.intValue(); // null mI -> g connected
pw.println((best_node_id + 1));
}
}
continue;
} else if (gj.getNumNodes() <= 2) {
++_totActiveNodes;
++_totLeafNodes;
// figure out max. node-weight
int best_node_id = -1;
double maxw = Double.NEGATIVE_INFINITY;
for (int m = 0; m < gj.getNumNodes(); m++) {
Double nmwD = gj.getNode(m).getWeightValue("value");
double nmw = nmwD == null ? 1.0 : nmwD.doubleValue();
if (nmw > maxw) {
maxw = nmw;
Integer mI = (Integer) gj.getNodeLabel(m);
best_node_id = mI == null ? m : mI.intValue(); // null mI -> g connected
}
}
_totActiveNodeWeights += maxw;
pw.println((best_node_id + 1));
continue;
}
// double bound = Double.MAX_VALUE;
double bound = 0;
String pdahost = "localhost";
if (params.containsKey("pdahost")) pdahost = (String) params.get("pdahost");
int pdaport = 7981;
if (params.containsKey("pdaport")) pdaport = ((Integer) params.get("pdaport")).intValue();
String cchost = "localhost";
if (params.containsKey("cchost")) cchost = (String) params.get("cchost");
int ccport = 7899;
if (params.containsKey("ccport")) ccport = ((Integer) params.get("ccport")).intValue();
String acchost = "localhost";
if (params.containsKey("acchost")) acchost = (String) params.get("acchost");
int accport = 7900;
if (params.containsKey("accport")) accport = ((Integer) params.get("accport")).intValue();
// DBBTree.init(g, bound, pdahost, pdaport, cchost, ccport, acchost, accport);
// DBBTree t = DBBTree.getInstance();
Boolean localSearchB = (Boolean) params.get("localsearch");
// if (localSearchB != null) t.setLocalSearch(localSearchB.booleanValue());
boolean localsearch = false;
if (localSearchB != null) localsearch = localSearchB.booleanValue();
AllChromosomeMakerClonableIntf maker =
(AllChromosomeMakerClonableIntf) params.get("localsearchtype");
// if (maker!=null) t.setLocalSearchType(maker);
Double ffD = (Double) params.get("ff");
/*
if (ffD!=null) {
DBBNode1.setFF(ffD.doubleValue());
DBBNode1.disallowFFChanges();
}
*/
double ff = 0.85;
if (ffD != null) ff = ffD.doubleValue();
Integer tlvlI = (Integer) params.get("tightenboundlevel");
/*
if (tlvlI != null && tlvlI.intValue() >= 1) t.setTightenUpperBoundLvl(
tlvlI.intValue());
*/
int tlvl = Integer.MAX_VALUE;
if (tlvlI != null && tlvlI.intValue() >= 1) tlvl = tlvlI.intValue();
Boolean usemaxsubsetsB = (Boolean) params.get("usemaxsubsets");
boolean usemaxsubsets = true;
if (usemaxsubsetsB != null)
// t.setUseMaxSubsets(usemaxsubsetsB.booleanValue());
usemaxsubsets = usemaxsubsetsB.booleanValue();
int kmax = Integer.MAX_VALUE;
Integer kmaxI = (Integer) params.get("maxitersinGBNS2A");
if (kmaxI != null && kmaxI.intValue() > 0)
// t.setMaxAllowedItersInGBNS2A(kmaxI.intValue());
kmax = kmaxI.intValue();
Boolean sortmaxsubsetsB = (Boolean) params.get("sortmaxsubsets");
boolean sortmaxsubsets = false;
if (sortmaxsubsetsB != null)
// t.setSortBestCandsInGBNS2A(sortmaxsubsetsB.booleanValue());
sortmaxsubsets = sortmaxsubsetsB.booleanValue();
Double avgpercextranodes2addD = (Double) params.get("avgpercextranodes2add");
double apen2a = 0.0;
if (avgpercextranodes2addD != null)
// t.setAvgPercExtraNodes2Add(avgpercextranodes2addD.doubleValue());
apen2a = avgpercextranodes2addD.doubleValue();
Boolean useGWMIN24BN2AB = (Boolean) params.get("useGWMIN2criterion");
boolean ugwm2 = false;
if (useGWMIN24BN2AB != null)
// t.setUseGWMIN24BestNodes2Add(useGWMIN24BN2AB.booleanValue());
ugwm2 = useGWMIN24BN2AB.booleanValue();
Double expandlocalsearchfactorD = (Double) params.get("expandlocalsearchfactor");
double elsf = 1.0;
if (expandlocalsearchfactorD != null)
// t.setLocalSearchExpandFactor(expandlocalsearchfactorD.doubleValue());
elsf = expandlocalsearchfactorD.doubleValue();
double mkb = Double.NEGATIVE_INFINITY;
Double minknownboundD = (Double) params.get("minknownbound");
if (minknownboundD != null) // t.setMinKnownBound(minknownboundD.doubleValue());
mkb = minknownboundD.doubleValue();
int maxchildren = Integer.MAX_VALUE;
Integer maxchildrenI = (Integer) params.get("maxnodechildren");
if (maxchildrenI != null && maxchildrenI.intValue() > 0)
// t.setMaxChildrenNodesAllowed(maxchildrenI.intValue());
maxchildren = maxchildrenI.intValue();
DBBNodeComparatorIntf bbcomp = (DBBNodeComparatorIntf) params.get("dbbnodecomparator");
// if (bbcomp!=null) t.setDBBNodeComparator(bbcomp);
DBBTree.init(
g,
bound,
pdahost,
pdaport,
cchost,
ccport,
acchost,
accport,
localsearch,
maker,
ff,
tlvl,
usemaxsubsets,
kmax,
sortmaxsubsets,
apen2a,
ugwm2,
elsf,
mkb,
maxchildren,
bbcomp);
DBBTree t = DBBTree.getInstance();
t.run();
int orsoln[] = t.getSolution();
int tan = 0;
double tanw = 0.0;
for (int i = 0; i < orsoln.length; i++) {
if (orsoln[i] == 1) {
Integer miI = (Integer) gj.getNodeLabel(i);
int mi = (miI == null) ? i : miI.intValue(); // null miI -> g connected
// System.out.print( mi + " ");
pw.println((mi + 1));
++tan;
Double twD = gj.getNode(i).getWeightValue("value");
tanw += (twD == null ? 1.0 : twD.doubleValue());
}
}
// tanw == t.getBound()
System.err.println("Total BB-nodes=" + t.getCounter());
System.err.println("Total leaf BB-nodes=" + t.getTotLeafNodes());
_totActiveNodes += tan;
_totActiveNodeWeights += tanw;
_totLeafNodes += t.getTotLeafNodes();
System.err.println(
"Total active nodes so far: "
+ _totActiveNodes
+ " active node weights="
+ _totActiveNodeWeights
+ " total overall trees leaf BB nodes="
+ _totLeafNodes);
System.err.println(
"Total #DLS searched performed: "
+ t.getNumDLSPerformed()
+ " Total time spent on DLS: "
+ t.getTimeSpentOnDLS());
}
pw.flush();
pw.close();
long time = System.currentTimeMillis() - start;
System.out.println("Best Soln = " + _totActiveNodeWeights);
System.out.println("\nWall-clock Time (msecs): " + time);
System.out.println("Done.");
System.exit(0);
} catch (Exception e) {
e.printStackTrace();
System.exit(-1);
}
}