mirror of https://github.com/jlizier/jidt
Added unit tests for Kraskov TE auto embedding via Ragwitz criteria -- verifies Issue 38. Also adds further tests for AIS auto embedding.
This commit is contained in:
joseph.lizier
parent
90fd9fbe89
commit
b45ada1d94
|
|
@ -1,5 +1,7 @@
|
|||
package infodynamics.measures.continuous.kraskov;
|
||||
|
||||
import java.util.Arrays;
|
||||
|
||||
import infodynamics.measures.continuous.ActiveInfoStorageCalculator;
|
||||
import infodynamics.utils.ArrayFileReader;
|
||||
import infodynamics.utils.MatrixUtils;
|
||||
|
|
@ -35,6 +37,8 @@ public class ActiveInfoStorageTester extends TestCase {
|
|||
public void testAutoEmbeddingRagwitz() throws Exception {
|
||||
ArrayFileReader afr = new ArrayFileReader("demos/data/SFI-heartRate_breathVol_bloodOx.txt");
|
||||
double[][] data = afr.getDouble2DMatrix();
|
||||
// Select data points 2350:3550
|
||||
data = MatrixUtils.selectRows(data, 2349, 3550-2350+1);
|
||||
|
||||
ActiveInfoStorageCalculatorKraskov ais = new ActiveInfoStorageCalculatorKraskov();
|
||||
// ais.setDebug(true);
|
||||
|
|
@ -54,8 +58,8 @@ public class ActiveInfoStorageTester extends TestCase {
|
|||
" and tau=" + optimisedTau + " optimised over kNNs=" +
|
||||
ais.getProperty(ActiveInfoStorageCalculatorKraskov.PROP_RAGWITZ_NUM_NNS));
|
||||
|
||||
// Test that the answer was k=3, tau=1 for this data set
|
||||
assertEquals(3, optimisedK);
|
||||
// Test that the answer was k=2, tau=1 for this data set (k=3 for full data set)
|
||||
assertEquals(2, optimisedK);
|
||||
assertEquals(1, optimisedTau);
|
||||
// Test that kNNs are equal to that used by the MI calculator when we have not set this
|
||||
assertEquals(ais.getProperty(MutualInfoCalculatorMultiVariateKraskov.PROP_K),
|
||||
|
|
@ -77,13 +81,35 @@ public class ActiveInfoStorageTester extends TestCase {
|
|||
double aisManualParamSetting = ais.computeAverageLocalOfObservations();
|
||||
assertEquals(aisOptimisedSingleThread, aisManualParamSetting, 0.00000001);
|
||||
|
||||
// Test that it works if we supply a validity vector:
|
||||
ais = new ActiveInfoStorageCalculatorKraskov();
|
||||
ais.setProperty(ActiveInfoStorageCalculatorKraskov.PROP_AUTO_EMBED_METHOD,
|
||||
ActiveInfoStorageCalculatorKraskov.AUTO_EMBED_METHOD_RAGWITZ);
|
||||
ais.setProperty(ActiveInfoStorageCalculatorKraskov.PROP_K_SEARCH_MAX, "4");
|
||||
ais.setProperty(ActiveInfoStorageCalculatorKraskov.PROP_TAU_SEARCH_MAX, "2");
|
||||
ais.initialise();
|
||||
boolean[] validity = new boolean[data.length];
|
||||
Arrays.fill(validity, true);
|
||||
ais.setObservations(MatrixUtils.selectColumn(data, 0), validity);
|
||||
double aisWithValidity = ais.computeAverageLocalOfObservations();
|
||||
assertEquals(aisOptimisedSingleThread, aisWithValidity, 0.00000001);
|
||||
assertEquals(optimisedK, Integer.parseInt(ais.getProperty(ActiveInfoStorageCalculator.K_PROP_NAME)));
|
||||
assertEquals(optimisedTau, Integer.parseInt(ais.getProperty(ActiveInfoStorageCalculator.TAU_PROP_NAME)));
|
||||
|
||||
// Finally, test that we can use a different number of kNNs to the MI calculator
|
||||
ais = new ActiveInfoStorageCalculatorKraskov();
|
||||
ais.setProperty(ActiveInfoStorageCalculatorKraskov.PROP_RAGWITZ_NUM_NNS, "8");
|
||||
ais.initialise(optimisedK, optimisedTau);
|
||||
ais.setProperty(ActiveInfoStorageCalculatorKraskov.PROP_RAGWITZ_NUM_NNS, "10");
|
||||
ais.setProperty(ActiveInfoStorageCalculatorKraskov.PROP_AUTO_EMBED_METHOD,
|
||||
ActiveInfoStorageCalculatorKraskov.AUTO_EMBED_METHOD_RAGWITZ);
|
||||
ais.setProperty(ActiveInfoStorageCalculatorKraskov.PROP_K_SEARCH_MAX, "4");
|
||||
ais.setProperty(ActiveInfoStorageCalculatorKraskov.PROP_TAU_SEARCH_MAX, "4");
|
||||
ais.initialise();
|
||||
ais.setObservations(MatrixUtils.selectColumn(data, 0));
|
||||
ais.computeAverageLocalOfObservations();
|
||||
assertEquals(8, Integer.parseInt(ais.getProperty(ActiveInfoStorageCalculatorKraskov.PROP_RAGWITZ_NUM_NNS)));
|
||||
double differentNNResult = ais.computeAverageLocalOfObservations();
|
||||
assertEquals(10, Integer.parseInt(ais.getProperty(ActiveInfoStorageCalculatorKraskov.PROP_RAGWITZ_NUM_NNS)));
|
||||
System.out.printf("Confirmed that we can change the number of nearest neighbours for Ragwitz optimisation, using " +
|
||||
"10 neighbours we get k=%s, tau=%s, ais=%.3f\n", ais.getProperty(ActiveInfoStorageCalculator.K_PROP_NAME),
|
||||
ais.getProperty(ActiveInfoStorageCalculator.TAU_PROP_NAME), differentNNResult);
|
||||
}
|
||||
|
||||
}
|
||||
|
|
|
|||
|
|
@ -18,6 +18,8 @@
|
|||
|
||||
package infodynamics.measures.continuous.kraskov;
|
||||
|
||||
import java.util.Arrays;
|
||||
|
||||
import infodynamics.utils.ArrayFileReader;
|
||||
import infodynamics.utils.MatrixUtils;
|
||||
|
||||
|
|
@ -395,4 +397,132 @@ public class TransferEntropyTester
|
|||
System.out.printf(" %.5f\n", result);
|
||||
assertEquals(expectedFromTRENTOOL1to2_k1l1, result, 0.000001);
|
||||
}
|
||||
|
||||
public void testAutoEmbeddingRagwitz() throws Exception {
|
||||
ArrayFileReader afr = new ArrayFileReader("demos/data/SFI-heartRate_breathVol_bloodOx.txt");
|
||||
double[][] data = afr.getDouble2DMatrix();
|
||||
// Select data points 2350:3550
|
||||
data = MatrixUtils.selectRows(data, 2349, 3550-2350+1);
|
||||
|
||||
TransferEntropyCalculatorKraskov teCalc = new TransferEntropyCalculatorKraskov();
|
||||
// teCalc.setDebug(true);
|
||||
|
||||
// Use one thread to test first:
|
||||
teCalc.setProperty(ConditionalMutualInfoCalculatorMultiVariateKraskov.PROP_NUM_THREADS, "1");
|
||||
teCalc.setProperty(TransferEntropyCalculatorKraskov.PROP_AUTO_EMBED_METHOD,
|
||||
TransferEntropyCalculatorKraskov.AUTO_EMBED_METHOD_RAGWITZ); // Embed both source and target
|
||||
teCalc.setProperty(TransferEntropyCalculatorKraskov.PROP_K_SEARCH_MAX, "5");
|
||||
teCalc.setProperty(TransferEntropyCalculatorKraskov.PROP_TAU_SEARCH_MAX, "5");
|
||||
teCalc.initialise();
|
||||
teCalc.setObservations(MatrixUtils.selectColumn(data, 0), MatrixUtils.selectColumn(data, 1));
|
||||
int optimisedK = Integer.parseInt(teCalc.getProperty(TransferEntropyCalculatorKraskov.K_PROP_NAME));
|
||||
int optimisedKTau = Integer.parseInt(teCalc.getProperty(TransferEntropyCalculatorKraskov.K_TAU_PROP_NAME));
|
||||
int optimisedL = Integer.parseInt(teCalc.getProperty(TransferEntropyCalculatorKraskov.L_PROP_NAME));
|
||||
int optimisedLTau = Integer.parseInt(teCalc.getProperty(TransferEntropyCalculatorKraskov.L_TAU_PROP_NAME));
|
||||
double teOptimisedSingleThread = teCalc.computeAverageLocalOfObservations();
|
||||
System.out.println("TE was " + teOptimisedSingleThread + " for k=" + optimisedK +
|
||||
",k_tau=" + optimisedKTau + ",l=" + optimisedL + ",l_tau=" + optimisedLTau +
|
||||
" optimised over kNNs=" +
|
||||
teCalc.getProperty(TransferEntropyCalculatorKraskov.PROP_RAGWITZ_NUM_NNS));
|
||||
|
||||
// Test that the answer was k=5, k_tau=1, l=2, l_tau=1 for this data set
|
||||
// (I've not checked this anywhere else, just making sure our result stays stable)
|
||||
assertEquals(5, optimisedK);
|
||||
assertEquals(1, optimisedKTau);
|
||||
assertEquals(2, optimisedL);
|
||||
assertEquals(1, optimisedLTau);
|
||||
// Test that kNNs are equal to that used by the MI calculator when we have not set this
|
||||
assertEquals(teCalc.getProperty(ConditionalMutualInfoCalculatorMultiVariateKraskov.PROP_K),
|
||||
teCalc.getProperty(TransferEntropyCalculatorKraskov.PROP_RAGWITZ_NUM_NNS));
|
||||
|
||||
// Test that we get the same answer by a multi-threaded approach
|
||||
// Use one thread to test first:
|
||||
teCalc.setProperty(ConditionalMutualInfoCalculatorMultiVariateKraskov.PROP_NUM_THREADS,
|
||||
ConditionalMutualInfoCalculatorMultiVariateKraskov.USE_ALL_THREADS);
|
||||
teCalc.initialise();
|
||||
teCalc.setObservations(MatrixUtils.selectColumn(data, 0), MatrixUtils.selectColumn(data, 1));
|
||||
double teOptimisedMultiThread = teCalc.computeAverageLocalOfObservations();
|
||||
assertEquals(teOptimisedSingleThread, teOptimisedMultiThread, 0.00000001);
|
||||
System.out.println("Answer unchanged by multi-threading");
|
||||
|
||||
// Test that optimisation looks the same if source and target swapped
|
||||
teCalc.initialise();
|
||||
teCalc.setObservations(MatrixUtils.selectColumn(data, 1), MatrixUtils.selectColumn(data, 0));
|
||||
assertEquals(teOptimisedSingleThread, teOptimisedMultiThread, 0.00000001);
|
||||
int optimisedSwappedK = Integer.parseInt(teCalc.getProperty(TransferEntropyCalculatorKraskov.K_PROP_NAME));
|
||||
int optimisedSwappedKTau = Integer.parseInt(teCalc.getProperty(TransferEntropyCalculatorKraskov.K_TAU_PROP_NAME));
|
||||
int optimisedSwappedL = Integer.parseInt(teCalc.getProperty(TransferEntropyCalculatorKraskov.L_PROP_NAME));
|
||||
int optimisedSwappedLTau = Integer.parseInt(teCalc.getProperty(TransferEntropyCalculatorKraskov.L_TAU_PROP_NAME));
|
||||
assertEquals(optimisedK, optimisedSwappedL);
|
||||
assertEquals(optimisedKTau, optimisedSwappedLTau);
|
||||
assertEquals(optimisedL, optimisedSwappedK);
|
||||
assertEquals(optimisedLTau, optimisedSwappedKTau);
|
||||
System.out.println("Ragwitz auto-embedding for source and destination swaps around when we swap source and target");
|
||||
|
||||
// Test that we can turn optimisation off now and we can hard code the parameters:
|
||||
teCalc.setProperty(TransferEntropyCalculatorKraskov.PROP_AUTO_EMBED_METHOD,
|
||||
TransferEntropyCalculatorKraskov.AUTO_EMBED_METHOD_NONE); // No auto embedding
|
||||
teCalc.initialise(2, 2, 2, 2, 2);
|
||||
teCalc.setObservations(MatrixUtils.selectColumn(data, 0), MatrixUtils.selectColumn(data, 1));
|
||||
double teDifferentParams = teCalc.computeAverageLocalOfObservations();
|
||||
assertFalse(teDifferentParams == teOptimisedSingleThread);
|
||||
assertEquals(2, Integer.parseInt(teCalc.getProperty(TransferEntropyCalculatorKraskov.K_PROP_NAME)));
|
||||
assertEquals(2, Integer.parseInt(teCalc.getProperty(TransferEntropyCalculatorKraskov.K_TAU_PROP_NAME)));
|
||||
assertEquals(2, Integer.parseInt(teCalc.getProperty(TransferEntropyCalculatorKraskov.L_PROP_NAME)));
|
||||
assertEquals(2, Integer.parseInt(teCalc.getProperty(TransferEntropyCalculatorKraskov.L_TAU_PROP_NAME)));
|
||||
System.out.printf("Auto-embedding goes away when we request no auto embedding (result now TE(k=%d,k_tau=%d,l=%d,l_tau=%d,u=%d)=%.3f)\n",
|
||||
2, 2, 2, 2, 2, teDifferentParams);
|
||||
|
||||
// Test that we get the same answer by setting these parameters
|
||||
teCalc = new TransferEntropyCalculatorKraskov();
|
||||
teCalc.initialise(optimisedK, optimisedKTau, optimisedL, optimisedLTau, 1);
|
||||
teCalc.setObservations(MatrixUtils.selectColumn(data, 0), MatrixUtils.selectColumn(data, 1));
|
||||
double teManual = teCalc.computeAverageLocalOfObservations();
|
||||
assertEquals(teOptimisedSingleThread, teManual, 0.00000001);
|
||||
System.out.println("Result stable to hard-coding auto-embedded parameters");
|
||||
|
||||
// Test optimising destination only
|
||||
teCalc.initialise();
|
||||
// auto embed destination only
|
||||
teCalc.setProperty(TransferEntropyCalculatorKraskov.PROP_AUTO_EMBED_METHOD,
|
||||
TransferEntropyCalculatorKraskov.AUTO_EMBED_METHOD_RAGWITZ_DEST_ONLY);
|
||||
teCalc.setProperty(TransferEntropyCalculatorKraskov.PROP_K_SEARCH_MAX, "5");
|
||||
teCalc.setProperty(TransferEntropyCalculatorKraskov.PROP_TAU_SEARCH_MAX, "5");
|
||||
// Explicitly set the source embedding params
|
||||
teCalc.setProperty(TransferEntropyCalculatorKraskov.L_PROP_NAME, "1");
|
||||
teCalc.setProperty(TransferEntropyCalculatorKraskov.L_TAU_PROP_NAME, "1");
|
||||
teCalc.setObservations(MatrixUtils.selectColumn(data, 0), MatrixUtils.selectColumn(data, 1));
|
||||
double teOptimisedDestOnly = teCalc.computeAverageLocalOfObservations();
|
||||
assertFalse(teOptimisedDestOnly == teOptimisedSingleThread);
|
||||
assertEquals(optimisedK, Integer.parseInt(teCalc.getProperty(TransferEntropyCalculatorKraskov.K_PROP_NAME)));
|
||||
assertEquals(optimisedKTau, Integer.parseInt(teCalc.getProperty(TransferEntropyCalculatorKraskov.K_TAU_PROP_NAME)));
|
||||
assertEquals(1, Integer.parseInt(teCalc.getProperty(TransferEntropyCalculatorKraskov.L_PROP_NAME)));
|
||||
assertEquals(1, Integer.parseInt(teCalc.getProperty(TransferEntropyCalculatorKraskov.L_TAU_PROP_NAME)));
|
||||
System.out.printf("Auto-embedding for dest only does not embed the source (result now TE(k=%d,k_tau=%d,l=%d,l_tau=%d,u=%d)=%.3f)\n",
|
||||
optimisedK, optimisedKTau, 1, 1, 1, teOptimisedDestOnly);
|
||||
|
||||
// Finally, test that we can use a different number of kNNs to the MI calculator
|
||||
teCalc.initialise();
|
||||
// auto embed source and destination:
|
||||
teCalc.setProperty(TransferEntropyCalculatorKraskov.PROP_AUTO_EMBED_METHOD,
|
||||
TransferEntropyCalculatorKraskov.AUTO_EMBED_METHOD_RAGWITZ);
|
||||
teCalc.setProperty(TransferEntropyCalculatorKraskov.PROP_RAGWITZ_NUM_NNS, "8");
|
||||
teCalc.setObservations(MatrixUtils.selectColumn(data, 0), MatrixUtils.selectColumn(data, 1));
|
||||
|
||||
// TODO what happens with method to select start and end times if we
|
||||
// have not selected embedding parameters yet???
|
||||
teCalc = new TransferEntropyCalculatorKraskov();
|
||||
teCalc.setProperty(TransferEntropyCalculatorKraskov.PROP_AUTO_EMBED_METHOD,
|
||||
TransferEntropyCalculatorKraskov.AUTO_EMBED_METHOD_RAGWITZ);
|
||||
teCalc.setProperty(TransferEntropyCalculatorKraskov.PROP_K_SEARCH_MAX, "5");
|
||||
teCalc.setProperty(TransferEntropyCalculatorKraskov.PROP_TAU_SEARCH_MAX, "5");
|
||||
teCalc.initialise();
|
||||
boolean[] validity = new boolean[data.length];
|
||||
Arrays.fill(validity, true);
|
||||
teCalc.setObservations(MatrixUtils.selectColumn(data, 0), MatrixUtils.selectColumn(data, 1),
|
||||
validity, validity);
|
||||
double teOptimisedWithValidity = teCalc.computeAverageLocalOfObservations();
|
||||
assertEquals(teOptimisedSingleThread, teOptimisedWithValidity, 0.00000001);
|
||||
System.out.println("Answer unchanged by setting validity");
|
||||
}
|
||||
}
|
||||
|
|
|
|||
Loading…
Reference in New Issue