situation of warship, decoy and torpedo, that is, whether the torpedo can discover the
warship. In this paper, the problem is transformed into a binary classification scenario
by machine learning, and the feasibility of ensemble learning in warship operational
assistant decision-making is studied.
Ensemble learning refers to using simple, trainable and differentiated classifiers to
achieve accurate classification decisions through integrated voting [1]. Since ensemble
learning was first proposed to solve classification problems, and most of the ensemble
learning algorithms are used to design classifiers, thus ensemble learning is also called
classifier ensemble or multi-classifier system and so on. Both the classification accuracy can be improved and the risk of relying on a single classifier can be reduced
significantly by means of classifier ensemble. Therefore, it has been widely used in
bionics, finance, politics, and medicine [2]. But is it feasible to apply the ensemble
learning algorithm which is superior in other fields directly to the operational assistant
decision-making? As the “no free lunch” theorem states [3], there are many ensemble
learning algorithms, but there is no universal algorithm suitable for all data sets. It is
necessary to study the adaptability of ensemble learning to warship operational
assistant decision-making.
2 Data Set Description
The decoy jamming effect data set is generated by off-line simulation. In the simulation
process, the relative situation of warship, torpedo, and decoy can be described as five
parameters: the distance between torpedo and warship, the distance between decoy and
warship, the azimuth angle between decoy and warship, the azimuth angle between
torpedo and warship and the azimuth angle between warship and torpedo, which are,
respectively, denoted as x 1 , x 2 , x 3 , x 4 , x 5 . In this way, each sample is represented with a
five-dimensional vector, i.e., x = (x 1 , x 2 , …, x 5 ). The label y of sample x is 1 or −1,
y = −1 means that in the situation corresponding to x, the torpedo cannot find the
warship and vice versa.
In the simulation, each feature takes 10 values at equal intervals, as a result, 10
5
samples are generated for the training and testing of the ensemble classifier.
3 Algorithms Description
3.1 Single and Ensemble Classification Algorithms
In order to verify the effectiveness of ensemble learning, it is necessary to compare the
ensemble classification algorithm with typical classification algorithm. Decision tree is
one of the most widely used algorithms in the field of data mining. Therefore, the
decision tree was selected for comparative experiments.
In terms of ensemble learning, the Bagging algorithm [4] was selected for comparison, which used the reduced error pruning tree (REPTree) [5] as the base classification algorithm. Consider a classification problem with a set X = {x 1 , x 2 , …, x C } of
class labels and feature space XR
n . Let a pool of base classifiers C = {C 1 , C 2 , …, C L },
290
X. Wei and X. Fan
warship. In this paper, the problem is transformed into a binary classification scenario
by machine learning, and the feasibility of ensemble learning in warship operational
assistant decision-making is studied.
Ensemble learning refers to using simple, trainable and differentiated classifiers to
achieve accurate classification decisions through integrated voting [1]. Since ensemble
learning was first proposed to solve classification problems, and most of the ensemble
learning algorithms are used to design classifiers, thus ensemble learning is also called
classifier ensemble or multi-classifier system and so on. Both the classification accuracy can be improved and the risk of relying on a single classifier can be reduced
significantly by means of classifier ensemble. Therefore, it has been widely used in
bionics, finance, politics, and medicine [2]. But is it feasible to apply the ensemble
learning algorithm which is superior in other fields directly to the operational assistant
decision-making? As the “no free lunch” theorem states [3], there are many ensemble
learning algorithms, but there is no universal algorithm suitable for all data sets. It is
necessary to study the adaptability of ensemble learning to warship operational
assistant decision-making.
2 Data Set Description
The decoy jamming effect data set is generated by off-line simulation. In the simulation
process, the relative situation of warship, torpedo, and decoy can be described as five
parameters: the distance between torpedo and warship, the distance between decoy and
warship, the azimuth angle between decoy and warship, the azimuth angle between
torpedo and warship and the azimuth angle between warship and torpedo, which are,
respectively, denoted as x 1 , x 2 , x 3 , x 4 , x 5 . In this way, each sample is represented with a
five-dimensional vector, i.e., x = (x 1 , x 2 , …, x 5 ). The label y of sample x is 1 or −1,
y = −1 means that in the situation corresponding to x, the torpedo cannot find the
warship and vice versa.
In the simulation, each feature takes 10 values at equal intervals, as a result, 10
5
samples are generated for the training and testing of the ensemble classifier.
3 Algorithms Description
3.1 Single and Ensemble Classification Algorithms
In order to verify the effectiveness of ensemble learning, it is necessary to compare the
ensemble classification algorithm with typical classification algorithm. Decision tree is
one of the most widely used algorithms in the field of data mining. Therefore, the
decision tree was selected for comparative experiments.
In terms of ensemble learning, the Bagging algorithm [4] was selected for comparison, which used the reduced error pruning tree (REPTree) [5] as the base classification algorithm. Consider a classification problem with a set X = {x 1 , x 2 , …, x C } of
class labels and feature space XR
n . Let a pool of base classifiers C = {C 1 , C 2 , …, C L },
290
X. Wei and X. Fan