KR102188431B1 - Methods and apparatuses for selecting feature based on monte carlo tree search - Google Patents

Abstract

The present invention relates to a property selection method and apparatus based on a Monte Carlo tree search, and the property selection method based on a Monte Carlo tree search according to an embodiment of the present invention includes a feature selection tree consisting of nodes representing properties. Selecting a node using the maximum reward value of the node until a node that has not been expanded in the currently expanded attribute selection tree is reached for retrieval; Creating a lower node of the selected node to expand the tree; Expanding and simulating the remaining unexpanded nodes from the generated lower nodes to calculate an optimal subset of attributes; And updating the attribute selection tree by back propagating the simulation result through the selected node.

Description

Attribute selection method and device based on Monte Carlo tree search {METHODS AND APPARATUSES FOR SELECTING FEATURE BASED ON MONTE CARLO TREE SEARCH}

The present invention relates to a method and apparatus for selecting attributes based on a Monte Carlo tree search.

Conventionally, various approaches have been developed to solve the problem of feature selection. Depending on the search strategy used, they are mainly classified as exhaustive, heuristic or meta-heuristic. Applying an exhaustive search over data sets with many attributes is practically impossible because of the complexity. Heuristic approaches such as SFS (Sequential Forward Selection) and SBS (Sequential Backward Selection) suffer from the problem of nesting effect. It is also expensive to calculate.

Recently, meta-heuristic approaches such as GA (Genetic Algorithm), ACO (Ant Colony Optimization), PSO (Particle Swarm Optimization), multi-purpose evolution algorithm, and bat algorithm are attracting attention. However, they are in their infancy and use too many hyperparameters and tune the model, making the optimized performance too complex. Therefore, there is room for vast improvement, and new algorithms are greatly needed to overcome these problems. There is a great need for new algorithms that can efficiently achieve high performance with less model complexity.

Embodiments of the present invention use Monte Carlo Tree Selection (MCTS)-based attribute selection for the exploration and development of the entire attribute space for efficient classification of data, thereby finding an optimal subset of attributes. To provide a method and apparatus for selecting attributes based on tree search.

According to an embodiment of the present invention, in a Monte Carlo tree search-based attribute selection method performed by an attribute selection device, in order to search a feature selection tree consisting of nodes representing attributes, the extended Selecting a node using the maximum reward value of the node until a node that has not been expanded in the attribute selection tree is reached; Creating a lower node of the selected node to expand the tree; Expanding and simulating the remaining unexpanded nodes from the generated lower nodes to calculate an optimal subset of attributes; And updating the attribute selection tree by performing back propagation of the simulation result through the selected node, and a method for selecting a property based on a Monte Carlo tree search may be provided.

In the step of selecting the node, the tree may be searched by adding a node representing an attribute in the attribute selection tree using a Monte Carlo Tree Search.

In the step of selecting the node, a node may be selected using a maximum reward value representing the maximum accuracy of the node, the number of times the node has been visited, and the number of times the upper node of the node has been visited.

The simulating step may be performed by randomly expanding the remaining nodes until reaching the remaining unexpanded nodes from the generated lower node.

The simulating step may or may not include randomly including a node representing an unexpanded attribute in the path from the generated lower node to an end node for the remaining unexpanded nodes.

In the simulating step, at least one attribute subset included in the path from the generated lower node to the remaining unexpanded node may be calculated.

In the simulating step, if the accuracy of using the current attribute subset among the calculated at least one attribute subset is higher than the accuracy calculated in the previous iteration process, the current attribute subset is determined as the optimal attribute subset. I can.

The attribute selection tree may be expanded by adding a node indicating an attribute to the root node starting from a root node in which an attribute is not selected.

Each node in the attribute selection tree may indicate that an attribute is selected or that an attribute is not selected.

Each node in the attribute selection tree may have a lower node indicating that the attribute is selected and a lower node different from that of the attribute not being selected.

On the other hand, according to another embodiment of the present invention, a memory for storing at least one program; And a processor connected to the memory, wherein the processor executes the at least one program to search for a feature selection tree consisting of nodes representing attributes, from the extended attribute selection tree The generation is to select a node using the maximum reward value of the node until it reaches the unexpanded node, expand the tree by creating a sub-node of the selected node, and calculate an optimal subset of attributes. A property selection device based on a Monte Carlo tree search will be provided, which expands and simulates the remaining unexpanded nodes from the sub-node and updates the property selection tree by back propagating the simulation result through the selected node. I can.

The processor may search a tree by adding a node representing an attribute in the attribute selection tree using a Monte Carlo Tree Search.

The processor may select a node using a maximum reward value representing the maximum accuracy of the node, the number of times the node is visited, and the number of times the node's upper node is visited.

The processor may perform simulation by randomly expanding the remaining nodes until reaching the remaining unexpanded nodes from the generated lower node.

The processor may or may not randomly include a node indicating an unexpanded attribute in the path from the generated lower node to an end node for the remaining unexpanded nodes.

The processor may calculate at least one subset of attributes included in a path from the generated lower node to the remaining unexpanded node.

The processor may determine the current attribute subset as the optimal attribute subset if the accuracy of using the current attribute subset among the calculated at least one attribute subset is higher than the accuracy calculated in a previous iteration process. .

The attribute selection tree may be expanded by adding a node indicating an attribute to the root node starting from a root node in which an attribute is not selected.

Each node in the attribute selection tree may indicate that an attribute is selected or that an attribute is not selected.

The processor may have a lower node indicating that the attribute is selected and a lower node different from that of the attribute not being selected.

Meanwhile, according to another embodiment of the present invention, as a non-transitory computer-readable storage medium including at least one program executable by a processor, when the at least one program is executed by the processor, the processor causes: To search a feature selection tree consisting of nodes representing attributes, select a node using the node's maximum reward value until it reaches a node that has not been expanded in the currently expanded attribute selection tree. And, to generate a lower node of the selected node to expand the tree, and to calculate the optimal attribute subset, expand the remaining nodes that are not expanded from the generated lower node to simulate, and the simulation result to the selected node A non-transitory computer-readable storage medium including instructions for updating the attribute selection tree by back propagation may be provided.

Embodiments of the present invention may find an optimal subset of attributes by using attribute selection based on a Monte Carlo Tree Selection (MCTS) search for exploration and development of the entire attribute space for efficient classification of data.

Embodiments of the present invention can efficiently search the attribute space and find the best attribute subset within the least iteration. As an example, since only two hyper parameters need to be adjusted in the embodiments of the present invention, the embodiments of the present invention can be simple and flexible to handle compared to other general meta-heuristic methods.

1 is a diagram for describing an operation of a property selection device based on a Monte Carlo tree search according to an embodiment of the present invention.
FIG. 2 is a diagram for describing an attribute selection tree and a search process based on a Monte Carlo tree search according to an embodiment of the present invention.
3 is a flowchart illustrating a method of selecting an attribute based on a Monte Carlo tree search according to an embodiment of the present invention.
4 is a flowchart illustrating a method of selecting an attribute based on a Monte Carlo tree search according to an embodiment of the present invention.
5 is a block diagram illustrating a configuration of an attribute selection device based on a Monte Carlo tree search according to an embodiment of the present invention.
6 is a diagram illustrating a dataset used for verification and comparison purposes according to an embodiment of the present invention.
7 and 8 are diagrams for explaining a comparison result between an embodiment of the present invention and other approaches.
9 is a view for explaining the reduction in dimensionality achieved by an embodiment of the present invention.

In the present invention, various modifications may be made and various embodiments may be provided, and specific embodiments will be illustrated in the drawings and described in detail.

However, this is not intended to limit the present invention to a specific embodiment, it is to be understood to include all changes, equivalents, or substitutes included in the spirit and scope of the present invention.

Terms such as first and second may be used to describe various components, but the components should not be limited by the terms. These terms are used only for the purpose of distinguishing one component from another component. For example, without departing from the scope of the present invention, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element. The term and/or includes a combination of a plurality of related listed items or any of a plurality of related listed items.

When a component is referred to as being "connected" or "connected" to another component, it is understood that it may be directly connected or connected to the other component, but other components may exist in the middle. Should be. On the other hand, when a component is referred to as being "directly connected" or "directly connected" to another component, it should be understood that there is no other component in the middle.

The terms used in the present application are only used to describe specific embodiments, and are not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In the present application, terms such as "comprise" or "have" are intended to designate the presence of features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, but one or more other features. It is to be understood that the presence or addition of elements or numbers, steps, actions, components, parts, or combinations thereof, does not preclude in advance.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs. Terms as defined in a commonly used dictionary should be interpreted as having a meaning consistent with the meaning of the related technology, and should not be interpreted as an ideal or excessively formal meaning unless explicitly defined in this application. Does not.

Hereinafter, preferred embodiments of the present invention will be described in more detail with reference to the accompanying drawings. In describing the present invention, in order to facilitate an overall understanding, the same reference numerals are used for the same elements in the drawings, and duplicate descriptions for the same elements are omitted.

1 is a diagram for describing an operation of a property selection device based on a Monte Carlo tree search according to an embodiment of the present invention.

First, the attribute selection apparatus 100 based on a Monte Carlo tree search according to an embodiment of the present invention performs an attribute selection method based on a Monte Carlo tree search. In an embodiment of the present invention, attribute selection is regarded as a preliminary step prior to machine learning or data mining operations. The attribute selection method removes redundant or irrelevant attributes to maintain accurate information and data structures as much as possible, thereby creating an efficient and accurate prediction model to find the optimal attribute.

The attribute selection device 100 based on a Monte Carlo tree search according to an embodiment of the present invention classifies the attribute using a Monte Carlo tree search (MCTS) within a wrapper framework, while classifying a new and efficient attribute. You can choose. The attribute selection apparatus 100 according to an embodiment of the present invention may search for the best attribute subset by combining the advantages of tree search and random sampling. The attribute selection apparatus 100 according to an embodiment of the present invention may provide a new attribute selection tree in which exploration and development are well balanced within a limited calculation area.

The Monte Carlo Tree Search based Feature Selection (MOTiFS) method according to an embodiment of the present invention includes the robustness, dynamicity, and wrapper method of a Monte Carlo tree search. It uses a meta-heuristic algorithm that combines the high accuracy of ). The attribute selection method (MOTiFS) uses a Monte Carlo tree search as a search strategy of the wrapper method. The attribute selection method can create an attribute selection tree by adding a node representing the attribute starting from the root node where no attribute is selected.

Every iteration can generate a feature subset according to the Tree policy and the Default policy. The tree policy not only utilizes the property space expanded from the previous iteration, but also can gradually expand the tree and add new properties. The default policy can induce randomness by randomly selecting an attribute from an unselected attribute set.

The attribute selection apparatus 100 based on a Monte Carlo tree search according to an embodiment of the present invention may generate an optimal attribute subset through a combination of the above two policies. The operation structure of the attribute selection apparatus 100 based on a Monte Carlo tree search according to an embodiment of the present invention is illustrated in FIG. 1.

As shown in FIG. 1, the Monte Carlo tree search-based attribute selection apparatus 100 according to an embodiment of the present invention receives an original attribute set (F _original ) and calculates an attribute subset (F _subset ), and a classifier An optimal subset of attributes (F _best ) can be calculated through (Classifier). The attribute selection apparatus 100 may generate a subset of attributes according to a selection step, an expansion step, and a simulation step during a simulation step for attribute selection based on a Monte Carlo tree search. Then, the attribute selection device 100 may check the accuracy of the attribute subset by applying the classification criteria. In addition, classification accuracy is currently used as a compensation for simulation for attribute selection based on Monte Carlo tree search. The attribute selection device 100 may backpropagate the classification accuracy to the root node.

The attribute selection apparatus 100 based on a Monte Carlo tree search according to an embodiment of the present invention may formulate an attribute selection problem as a new attribute selection tree. Also, the attribute selection device 100 may use a Monte Carlo tree search (MCTS) for exploration and development of an attribute space. In addition, the attribute selection apparatus 100 may use a classifier for calculating classification performance and compensation during each MCTS simulation.

In an embodiment of the present invention, the root node of the attribute selection tree is in a state in which no attribute is selected, and all nodes have two lower nodes.

와

의 두 개 하위노드를 가진다. 노드

는 속성

가 선택되었음을 의미하고,

는 속성

가 선택되지 않았음을 의미한다. The node at level i-1 in the attribute selection tree is

Wow

It has two subnodes of. Node

The property

Means is selected,

The property

Means that is not selected.

The attribute selection device 100 may find a path showing the highest performance by using a Monte Carlo tree search. The path includes nodes representing the selected attribute. The operation of the attribute selection apparatus 100 based on a Monte Carlo tree search according to an embodiment of the present invention can be divided into four steps, each step as shown in FIG. 2. The attribute selection device 100 performs four steps including a selection step, an expansion step, a simulation step, and a backpropagation step.

FIG. 2 is a diagram for describing an attribute selection tree and a search process based on a Monte Carlo tree search according to an embodiment of the present invention.

First, definition 1 below will be defined in an embodiment of the present invention.

에 대해, 속성 선택 트리는 다음 조건 1 및 2를 만족하는 트리이다 :Definition 1. Set of attributes

For, the attribute selection tree is a tree that satisfies the following conditions 1 and 2:

이며, 속성이 선택되지 않음을 의미한다.Condition 1: The root is

And means that the attribute is not selected.

와

가 있으며,

이다.Condition 2: Node v _i at level i-1 has two subnodes

Wow

There is,

to be.

와

는 해당 속성

를 속성 서브 세트에 포함시키거나 제외시키는 것을 나타낸다. 루트 노드에서 리프 중 하나까지의 경로는 속성 서브세트를 나타낸다. 따라서 본 발명의 일 실시 예는 최상의 보상 (정확도)을 주는 경로를 찾는 것이다. 본 발명의 일 실시 예에 따른 속성 선택 장치(100)는 트리 구성 및 순회를 위해 몬테카를로 트리 검색(MCTS, Monte Carlo Tree Search)를 사용하고 최고의 정확도를 갖는 최적의 경로를 선택할 수 있다. 최적의 경로에는 최고의 정확도를 갖는 속성 서브세트에 포함된 노드들이 포함되어 있다.Where, node

Wow

Is the corresponding property

Indicates to include or exclude from a subset of attributes. The path from the root node to one of the leaves represents a subset of attributes. Therefore, one embodiment of the present invention is to find a path that gives the best compensation (accuracy). The attribute selection apparatus 100 according to an embodiment of the present invention may use a Monte Carlo Tree Search (MCTS) for tree construction and traversal and select an optimal path having the highest accuracy. The optimal path contains nodes included in the subset of attributes with the highest accuracy.

First, a description will be given from the selection step.

The attribute selection apparatus 100 based on a Monte Carlo tree search according to an embodiment of the present invention receives an original attribute set F _original and performs a Monte Carlo tree search. The attribute selection device 100 selects one path showing the highest reward from the attribute selection tree extended to date. Here, the reward indicates the accuracy of selecting an attribute. The attributes in the selected path are included in the attribute subset of the current iteration process.

The attribute selection device 100 determines a node to be selected at each level using an upper confidence bounds for trees (UCT) algorithm.

를 선택하면, 속성

는 현재의 속성 세트에 포함될 수 있다. 속성

가 선택되었다는 것은,

를 포함하는 속성 세트가 높은 리워드(reward)를 가져온다는 것을 의미한다. The attribute selection device 100 is a node at level i according to the UCT algorithm.

If you select, properties

May be included in the current attribute set. property

Is selected,

It means that a set of attributes containing a result of a high reward.

를 선택하면, 속성

는 포함되지 않는 것을 나타낸다. 여기서, 일반적인 UCT 알고리즘은 몬테카를로 트리 검색의 노드 선택 단계에서 일반적으로 사용되는 알고리즘으로 하기의 [수학식 1]과 같다. Attribute selection device 100 according to the UCT algorithm

If you select, properties

Indicates not included. Here, the general UCT algorithm is an algorithm commonly used in the node selection step of a Monte Carlo tree search and is as shown in [Equation 1] below.

와

는 노드 v와 v의 상위 노드 p가 방문된 횟수임.

는 노드 v에서 승리(게임이론의 관점에서) 횟수이다. 일반적인 UCT 알고리즘은 노드가 방문된 횟수로 나누어 각 노드의 리워드(reward)를 계산한다. here,

Wow

Is the number of times nodes v and v's parent node p have been visited.

Is the number of wins (in terms of game theory) at node v. A general UCT algorithm calculates the reward of each node by dividing it by the number of times the node is visited.

A general UCT algorithm is suitable from the point of view of a game theory where the reward is 1 (win) and 0 (loss). From the point of view of game theory, the purpose of the UCT algorithm is to find the node with the least visits and the greatest win.

In contrast, the attribute selection apparatus 100 based on a Monte Carlo tree search according to an embodiment of the present invention may find a path having a reward indicating maximum accuracy through attribute selection. Therefore, the attribute selection apparatus 100 based on Monte Carlo tree search according to an embodiment of the present invention modifies the UCT algorithm, does not use the number of visits and the reward, but uses the maximum reward at each node. do. [Equation 2] below shows a modified UCT algorithm used in an embodiment of the present invention.

)는 노드

의 최대 리워드(reward) 값이고, C > 0 은 상수이다.

와

는 노드

와 상위 노드인

가 방문된 횟수를 나타낸다. Where max(

) Is the node

Is the maximum reward value, and C> 0 is a constant.

Wow

Is the node

And the parent node

Represents the number of visits.

Next, the expansion (Expansion) step will be described.

The attribute selection device 100 adds a lower node of the node selected in the selection step. Here, the lower node may be referred to as a child node. The addition of new sub-nodes is made based on the UCT algorithm.

이

보다 크면 하위 노드

을 추가한다. 이는 현재의 속성 서브세트(subset)에 속성

이 포함됨을 의미한다. 반대로 하위 노드

이 추가되면 속성

은 현재의 속성 서브세트(subset)에 포함되지 않는다.The attribute selection apparatus 100 based on a Monte Carlo tree search according to an embodiment of the present invention uses a UCT algorithm,

this

If greater than the child node

Add This is an attribute in the current subset of attributes.

This means that it is included. Reverse child node

When this property is added

Is not included in the current subset of attributes.

Next, a simulation step will be described.

라면,

까지의 경로(path)와 그의 하위 노드가 무작위로 선택된다. Includes randomness in generating attribute subsets. That is, the attribute selection apparatus 100 based on a Monte Carlo tree search according to an embodiment of the present invention randomly includes or does not include unexpanded attributes. The most recently expanded node

Ramen,

The path to and its subnodes are randomly selected.

로부터 속성

까지의 포함은 선택 단계와 확장 단계에서 결정된다. 속성

로부터 속성

까지의 포함은 시뮬레이션 단계에서 무작위로 결정될 수 있다. In the current subset of attributes, the attribute

Property from

The inclusion of up to is determined in the selection and expansion steps. property

Property from

The inclusion to can be determined at random during the simulation phase.

The attribute selection apparatus 100 based on a Monte Carlo tree search according to an embodiment of the present invention undergoes a selection step, an expansion step, and a simulation step through a tree search and a random search, and an optimal attribute subset is performed. (subset) can be determined.

Monte Carlo tree search based on attribute selection unit 100 in accordance with one embodiment of the present invention as shown in [Equation 3] below, by simulating the accuracy of the classifier (Classifier) on the attribute subsets (F _subset) calculated The optimal subset of attributes (F _best ) can be calculated.

는 분류기의 정확도를 의미한다. 현재의 속성 서브세트(subset)을 이용했을 때의 정확도가 이전의 정확도보다 높으면, 현재의 속성 서브세트가 최적의 속성 서브세트로 결정될 수 있다. Here, Q _simulation is a simulation reward,

Means the accuracy of the classifier. If the accuracy when using the current attribute subset is higher than the previous accuracy, the current attribute subset may be determined as the optimal attribute subset.

3 is a flowchart illustrating a method of selecting an attribute based on a Monte Carlo tree search according to an embodiment of the present invention.

In step S101, the attribute selection apparatus 100 based on a Monte Carlo tree search according to an embodiment of the present invention receives an original attribute set.

In step S102, the attribute selection apparatus 100 selects a node by using the maximum reward value of the node.

In step S103, the attribute selection apparatus 100 expands the tree by creating a lower node of the node.

In step S104, the attribute selection device 100 expands and simulates the remaining nodes that are not expanded from the lower nodes.

In step S105, the attribute selection device 100 calculates an attribute subset through simulation.

In step S106, the attribute selection device 100 checks whether the accuracy of the calculated attribute subset exceeds the previously calculated accuracy.

In step S107, if the accuracy of the calculated attribute subset does not exceed the previously calculated accuracy, the attribute selection tree is updated by backpropagating the simulation result. The attribute selection device 100 may repeatedly perform step S107 according to the confirmation result in step S106.

In step S108, when the accuracy of the calculated attribute subset exceeds the previously calculated accuracy, the attribute selection apparatus 100 calculates an optimal attribute subset having the highest reward.

4 is a flowchart illustrating a method of selecting an attribute based on a Monte Carlo tree search according to an embodiment of the present invention.

In step S201, the Monte Carlo tree search-based attribute selection apparatus 100 according to an embodiment of the present invention includes an original attribute set F = (f ₁ , f ₂ , ... , f _n } is input.

In step S202, the attribute selection device 100 performs a selection step. The attribute selection device 100 searches the previously expanded attribute selection tree until a node having a non-terminal and unexpanded child node is reached. In addition, the attribute selection device 100 selects a node having the highest score calculated using [Equation 2] according to the modified UCT algorithm at each tree level.

In step S203, the attribute selection device 100 performs an expansion step. The attribute selection device 100 adds a new child node to an urgent node (the node selected last during the selection step). The attribute selection device 100 adds a child node having the highest score calculated using an equation according to the modified UCT algorithm.

In step S204, the attribute selection device 100 performs a simulation step. Here, the attribute may be selected from the remaining unexpanded attributes with a uniform probability of being selected or not selected.

In step S205, the attribute selection device 100 calculates an attribute subset as shown in [Equation 4] below.

Here, a(v _i ) represents an operation (f _i+1 or φ _i+1 ) taken at node v _i .

In step S206, the attribute selection device 100 calculates a simulation reward as in [Equation 3] above.

In step S207, the attribute selection device 100 performs a backpropagation step. After performing step S206, the attribute selection device 100 repeatedly updates the current active path until a Stopping Criteria is satisfied.

Thereafter, in step S208, the attribute selection device 100 calculates an optimal attribute subset. The optimal attribute subset means the attribute subset with the highest reward.

5 is a block diagram illustrating a configuration of an attribute selection device based on a Monte Carlo tree search according to an embodiment of the present invention.

As shown in FIG. 5, a Monte Carlo tree search-based attribute selection apparatus 100 according to an embodiment of the present invention includes a memory 110 and a processor 120. However, not all of the illustrated components are essential components. The attribute selection device 100 may be implemented by more components than the illustrated components, or the property selection device 100 may be implemented by fewer components.

Hereinafter, a detailed configuration and operation of each component of the attribute selection apparatus 100 based on the Monte Carlo tree search of FIG. 5 will be described.

The memory 110 stores at least one program.

The processor 120 is connected to the memory 110. Processor 120, by executing at least one program, in order to search for a feature selection tree consisting of nodes representing attributes, until a node that has not been expanded in the attribute selection tree expanded to date is reached. Remaining unexpanded nodes from the created sub-nodes to select a node using the node's maximum reward value, create sub-nodes of that selected node to expand the tree, and yield the optimal subset of attributes. Expand and simulate, and update the attribute selection tree by back propagating the simulation result through the selected node.

According to embodiments, the processor 120 may search the tree by adding a node representing an attribute in the attribute selection tree using Monte Carlo Tree Search.

According to embodiments, the processor 120 may select a node using a maximum reward value representing the maximum accuracy of the node, the number of times the node is visited, and the number of times the node's upper node is visited.

According to embodiments, the processor 120 may perform simulation by randomly expanding the remaining nodes from the generated lower node until reaching the remaining unexpanded nodes.

According to embodiments, the processor 120 may or may not include, with respect to the remaining unexpanded nodes, randomly including a node representing an unexpanded attribute in the path from the generated lower node to the end node.

According to embodiments, the processor 120 may calculate at least one subset of attributes included in a path from a generated lower node to a remaining node that has not been expanded.

According to embodiments, if the accuracy of using the current attribute subset among the calculated at least one attribute subset is higher than the accuracy calculated in the previous iteration process, the processor 120 determines the current attribute subset as the optimal attribute. It can be determined by a subset.

According to embodiments, the attribute selection tree may be expanded by starting from a root node in which an attribute is not selected, and adding a node indicating an attribute to the root node.

According to embodiments, each node in the attribute selection tree may indicate that an attribute is selected or that the attribute is not selected.

According to embodiments, the processor 120 may have a lower node indicating that the attribute is selected and a lower node different from that of the attribute not being selected.

6 is a diagram illustrating a dataset used for verification and comparison purposes according to an embodiment of the present invention.

As shown in FIG. 6, 25 benchmark data sets were used for verification and comparison purposes according to an embodiment of the present invention. To do this, we pulled 24 data sets from two previously published repositories. However, the "Klekota Roth fingerprint (KRFP)" data set representing the fingerprint of chemical compounds was taken from the 5-HT5A data set to classify active or inactive compounds.

7 and 8 are diagrams for explaining a comparison result between an embodiment of the present invention and other approaches.

Data between a Monte Carlo Tree Search Based Feature Selection (MOTiFS) representing a Monte Carlo search-based attribute selection method according to an embodiment of the present invention, and SFSW, SFS, SBS, FS-FS, DEMOFS, BA, and PSO as other methods The comparison result for each set is shown in FIG. 7. 7 shows the average accuracy (Avg. ACC.) and the number of selection attributes (# Sel. Feat.) according to 5-NN.

In FIG. 8, a comparison result for each data set between a Monte Carlo Tree Search Based Feature Selection (MOTiFS) representing a Monte Carlo search-based attribute selection method according to an embodiment of the present invention and other methods SCE, CCE, and PDE-2 are shown in FIG. Is shown. 8 shows the average accuracy (Avg. ACC.) and the number of selection attributes (# Sel. Feat.) according to 3-NN.

As described above, it can be seen that the Monte Carlo search-based attribute selection method (MOTiFS) according to an embodiment of the present invention has superior performance in most data sets compared to other approaches.

9 is a view for explaining the reduction in dimensionality achieved by an embodiment of the present invention.

9 provides a graphical representation of dimensional reduction (DR) achieved by a Monte Carlo search-based attribute selection method (MOTiFS) according to an embodiment of the present invention. The Monte Carlo search-based attribute selection method (MOTiFS) outperformed all other methods in terms of accuracy and dimension (DR) of all data sets. In "Ionosphere", "Arrhythmia", "Multiple features", "Sonar", "Musk 1", "Splice", "KRFP" and "Libras movement", the Monte Carlo search-based property selection method (MOTiFS) is dimensional reduction ( DR) value is obtained over 50%, showing higher accuracy than other methods.

The Monte Carlo tree search-based attribute selection method according to the embodiments of the present invention described above may be implemented as a computer-readable code on a computer-readable recording medium. The Monte Carlo tree search-based attribute selection method according to embodiments of the present invention may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer-readable recording medium.

A non-transitory computer-readable storage medium including at least one program executable by a processor, wherein when the at least one program is executed by the processor, the processor causes: an attribute selection tree consisting of nodes representing attributes. selection tree), select a node using the maximum reward value of the node until it reaches a node that has not been expanded in the attribute selection tree that has been expanded so far, and creates a sub-node of the selected node. To expand the tree and calculate the optimal attribute subset, the generated lower node is simulated by expanding the remaining unexpanded nodes, and the simulation result is back propagated through the selected node to select the attribute A non-transitory computer-readable storage medium may be provided containing instructions for causing a tree to be updated.

The above-described method according to the present invention may be implemented as a computer-readable code on a computer-readable recording medium. The computer-readable recording medium includes all kinds of recording media in which data that can be decoded by a computer system is stored. For example, there may be read only memory (ROM), random access memory (RAM), magnetic tape, magnetic disk, flash memory, optical data storage device, and the like. In addition, the computer-readable recording medium can be distributed to a computer system connected through a computer communication network, and stored and executed as code that can be read in a distributed manner.

Although described above with reference to the drawings and examples, it does not mean that the protection scope of the present invention is limited by the drawings or examples, and those skilled in the art It will be appreciated that various modifications and changes can be made to the present invention without departing from the spirit and scope.

Specifically, the described features may be implemented in digital electronic circuitry, or computer hardware, firmware, or combinations thereof. Features may be executed in a computer program product implemented in storage in a machine-readable storage device, for example, for execution by a programmable processor. And the features can be performed by a programmable processor executing a program of directives to perform the functions of the described embodiments by operating on input data and generating output. The described features include at least one programmable processor, at least one input device, and at least one output coupled to receive data and directives from the data storage system and to transmit data and directives to the data storage system. It can be executed within one or more computer programs that can be executed on a programmable system including the device. A computer program includes a set of directives that can be used directly or indirectly within a computer to perform a specific action on a given result. A computer program is written in any form of a programming language, including compiled or interpreted languages, and is included as a module, element, subroutine, or other unit suitable for use in another computer environment, or as a independently operable program. It can be used in any form.

Suitable processors for execution of a program of directives include, for example, both general and special purpose microprocessors, and either a single processor or multiple processors of a different type of computer. Storage devices suitable for implementing computer program directives and data implementing the described features are, for example, semiconductor memory devices such as EPROM, EEPROM, and flash memory devices, magnetic devices such as internal hard disks and removable disks. Devices, magneto-optical disks, and all types of non-volatile memory including CD-ROM and DVD-ROM disks. The processor and memory may be integrated within application-specific integrated circuits (ASICs) or added by ASICs.

The present invention described above is described on the basis of a series of functional blocks, but is not limited by the above-described embodiments and the accompanying drawings, and various substitutions, modifications and changes within the scope not departing from the technical spirit of the present invention It will be apparent to those of ordinary skill in the art to which this invention pertains.

Combinations of the above-described embodiments are not limited to the above-described embodiments, and various types of combinations as well as the above-described embodiments may be provided according to implementation and/or need.

In the above-described embodiments, the methods are described on the basis of a flowchart as a series of steps or blocks, but the present invention is not limited to the order of steps, and certain steps may occur in a different order or concurrently with those described above. have. In addition, those of ordinary skill in the art understand that the steps shown in the flowchart are not exclusive, other steps are included, or one or more steps in the flowchart may be deleted without affecting the scope of the present invention. You can understand.

The above-described embodiments include examples of various aspects. Although not all possible combinations for representing the various aspects can be described, those of ordinary skill in the art will recognize that other combinations are possible. Accordingly, the present invention will be said to cover all other replacements, modifications and changes falling within the scope of the following claims.

100: property selection device
110: memory
120: processor

Claims (21)

In a Monte Carlo tree search-based attribute selection method performed by an attribute selection device,
To search a feature selection tree consisting of nodes representing attributes, select a node using the node's maximum reward value until it reaches a node that has not been expanded in the currently expanded attribute selection tree. Step to do;
Creating a lower node of the selected node to expand the tree;
Expanding and simulating the remaining unexpanded nodes from the generated lower nodes to calculate an optimal subset of attributes; And
And updating the attribute selection tree by back propagating the simulation result through the selected node,
The simulating step calculates at least one attribute subset included in a path from the generated lower node to the remaining unexpanded node, a Monte Carlo tree search-based attribute selection method. The method of claim 1,
The step of selecting the node,
A Monte Carlo tree search-based attribute selection method for searching a tree by adding a node representing an attribute in the attribute selection tree using Monte Carlo Tree Search. The method of claim 1,
The step of selecting the node,
A method for selecting a node based on a Monte Carlo tree search, wherein a node is selected using a maximum reward value representing the maximum accuracy of the node, the number of times the node is visited, and the number of times the upper node of the node is visited. The method of claim 1,
The simulating step,
A method for selecting a property based on a Monte Carlo tree search by randomly expanding and simulating the remaining nodes from the generated lower node to the remaining unexpanded nodes. The method of claim 1,
The simulating step,
For the remaining non-expanded nodes, a node representing an unexpanded attribute is randomly included or not included in a path from the generated lower node to an end node, based on a Monte Carlo tree search method. delete The method of claim 1,
The simulating step,
If the accuracy of using the current attribute subset among the calculated at least one attribute subset is higher than the accuracy calculated in the previous iteration process, the current attribute subset is determined as the optimal attribute subset, based on a Monte Carlo tree search How to choose properties of the. The method of claim 1,
The attribute selection tree is a Monte Carlo tree search-based attribute selection method in which a node representing an attribute is added to the root node and expanded starting from a root node in which no attribute is selected. The method of claim 1,
Each node in the attribute selection tree indicates that an attribute is selected or an attribute is not selected. A method for selecting an attribute based on a Monte Carlo tree search. The method of claim 1,
Each node in the attribute selection tree,
A Monte Carlo tree search-based attribute selection method having a sub-node indicating that the attribute is selected and another sub-node indicating that the attribute is not selected. A memory for storing at least one program; And
Including a processor connected to the memory,
The processor, by executing the at least one program,
To search a feature selection tree consisting of nodes representing attributes, select a node using the node's maximum reward value until it reaches a node that has not been expanded in the currently expanded attribute selection tree. and,
Create a sub-node of the selected node to expand the tree,
In order to calculate an optimal attribute subset, the generated lower node is simulated by expanding the remaining unexpanded nodes, but at least one attribute subset included in the path from the generated lower node to the remaining unexpanded nodes is Calculate,
A property selection device based on a Monte Carlo tree search for updating the property selection tree by back propagating the simulation result through the selected node. The method of claim 11,
The processor,
A property selection device based on a Monte Carlo tree search for searching a tree by adding a node representing a property in the property selection tree using a Monte Carlo Tree Search. The method of claim 11,
The processor,
An attribute selection device based on a Monte Carlo tree search for selecting a node using a maximum reward value representing the maximum accuracy of the node, the number of times the node is visited, and the number of times the upper node of the node is visited. The method of claim 11,
The processor,
A property selection device based on a Monte Carlo tree search, which randomly expands and simulates the remaining nodes from the generated lower node to the non-expanded remaining nodes. The method of claim 11,
The processor,
For the remaining unexpanded nodes, a node representing an unexpanded attribute is randomly included or not included in a path from the generated lower node to an end node, based on a Monte Carlo tree search. delete The method of claim 11,
The processor,
If the accuracy of using the current attribute subset among the calculated at least one attribute subset is higher than the accuracy calculated in the previous iteration process, the current attribute subset is determined as the optimal attribute subset, based on a Monte Carlo tree search Property selection device. The method of claim 11,
The attribute selection tree is a Monte Carlo tree search-based attribute selection device in which a node representing an attribute is added to the root node and expanded starting from a root node in which no attribute is selected. The method of claim 11,
Each node in the attribute selection tree indicates that an attribute is selected or an attribute is not selected. A Monte Carlo tree search-based attribute selection device. The method of claim 11,
The processor,
An attribute selection device based on a Monte Carlo tree search, having a lower node indicating that the attribute is selected and another sub node indicating that the attribute is not selected. A non-transitory computer-readable storage medium comprising at least one program executable by a processor, wherein the at least one program, when executed by the processor, causes the processor to:
To search a feature selection tree consisting of nodes representing attributes, select a node using the node's maximum reward value until it reaches a node that has not been expanded in the currently expanded attribute selection tree. and,
Create a sub-node of the selected node to expand the tree,
In order to calculate an optimal attribute subset, the generated lower node is simulated by expanding the remaining unexpanded nodes, but at least one attribute subset included in the path from the generated lower node to the remaining unexpanded nodes is Calculate,
A non-transitory computer-readable storage medium comprising instructions for causing the simulation result to be back propagated through the selected node to update the attribute selection tree.

Images