System and method for placement with minimal dead space

Abstract

A dead space minimization placement system and method are disclosed. The present invention can improve space utilization efficiency by generating an occupancy map, predicting dead space, analyzing connection importance, controlling entropy, and providing feedback through a policy neural network based on a layout in which a placement target object is placed.

Description

Technology Field

[0001] The present invention relates to a dead space minimization placement system and method, and more specifically, to a dead space minimization placement system and method that improves space utilization efficiency by performing occupancy map generation, dead space prediction, connection importance analysis, entropy control, and feedback provision through a policy neural network based on a layout in which a target object is placed. Background Technology

[0002] Dead space refers to inefficient space that is not utilized between circuit layouts or wiring in semiconductor design, but this concept is also gaining attention in various industrial fields.

[0003] In the aerospace field, radiation shielding design is essential when applying commercial off-the-shelf (COTS) semiconductors to the space environment; in this process, dead space can be utilized to efficiently deploy shielding materials or contribute to thermal control and structural stability.

[0004] Furthermore, in the design of medical devices or industrial equipment, space efficiency can be improved by reducing or utilizing dead space, and functional designs such as heat dissipation or electromagnetic shielding can be optimized.

[0005] Physical placement is a key process for efficiently placing circuits or other objects to be placed into a limited two-dimensional placement area.

[0006] Recently, the occurrence of dead space is being minimized by using automated placement algorithms trained through reinforcement learning to classify objects to be placed by size and place them sequentially.

[0007] FIG. 1 is an example diagram illustrating the placement results of a placement system according to the prior art.

[0008] As shown in FIG. 1, when placing objects to be placed in a limited two-dimensional placement area (10), indirect performance indicators such as length, spacing, and congestion have been used as targets for optimization.

[0009] However, this arrangement method leaves irregular gaps between design objects and does not account for dead space (20, 21), which is unused space within the actual arrangement area (10), resulting in inefficient space utilization. Prior art literature

[0010] Korean Registered Patent Publication No. 10-2634706 (Title of Invention: Integrated Circuit Design Apparatus and Method for Minimizing Dead Space, Registration Date: 2024.02.02.) The problem to be solved

[0011] To solve these problems, the present invention aims to provide a dead space minimization placement system and method that improves space utilization efficiency by performing occupancy map generation, dead space prediction, connection importance analysis, entropy control, and feedback provision through a policy neural network based on a layout in which a target object is placed. means of solving the problem

[0012] To achieve the above-mentioned objective, one embodiment of the present invention is a dead space minimization placement system, wherein when information related to a design object is input, the system generates an occupancy map by displaying the placement area occupied by the design object based on the current placement state in a two-dimensional placement area, receives the occupancy map through a policy network to predict the dead space that may occur after the placement of the design object and outputs a placement policy, wherein the system generates an embedding reflecting the importance between adjacent nodes through a Graph Attention Network (GAT) based on the connection relationship between the design objects and reflects it in the placement policy, executes the placement of the design object according to the placement policy, and reflects the occupancy state resulting from the placement execution in the occupancy map; characterized by being a placement system.

[0013] In addition, the deployment system according to the above embodiment is characterized by adjusting an entropy coefficient that controls the balance of exploratoryness and convergence in the action selection process of the policy network based on the entropy and performance indicators of the policy network.

[0014] In addition, the placement system according to the above embodiment is characterized by generating feedback information based on occupancy status and space efficiency information after the placement of the design object, and reflecting this in the learning process of the policy network.

[0015] In addition, the placement system according to the above embodiment is characterized by comprising: a occupancy map analysis unit that generates an occupancy map by displaying the placement area occupied by the design object based on the current placement state in a two-dimensional placement area when information related to the design object is input; a GAT analysis unit that generates an embedding reflecting the importance between adjacent nodes through a GAT (Graph Attention Networks) based on the connection relationship between the design objects and reflects it in the placement policy; a dead space policy decision unit that receives information related to the occupancy map from the occupancy map analysis unit through a policy network, predicts the dead space that may occur after the placement of the design object and outputs a placement policy, and operates so that the occupancy state resulting from the placement execution is reflected in the occupancy map; an entropy control unit that adjusts an entropy coefficient controlling the balance of exploratoryness and convergence in the action selection process of the policy network based on the entropy and performance indicators of the policy network; and a placement unit that executes the placement of the design object according to the placement policy.

[0016] In addition, the occupancy map analysis unit according to the above embodiment is characterized by comprising: an input interface unit that generates a graph structure composed of nodes and edges for the connection relationship between design objects based on design object information and initializes a placement area; an occupancy map generation unit that displays the occupancy area occupied by a design object in the two-dimensional placement area based on the position and size information of the currently placed design object and generates an updated occupancy map according to the position and direction of the newly placed design object; and an occupancy information visualization unit that converts the generated occupancy map into arbitrary visualization information and outputs it.

[0017] In addition, the GAT analysis unit according to the above embodiment is characterized by comprising: a GAT network unit that performs node embedding by reflecting weights based on importance with respect to neighbor nodes, based on a graph structure composed of nodes and edges; a node embedding unit that generates an embedding vector for each node according to attributes including the size, type, and unique information of the design object; and a weight evaluation unit that calculates an attention score according to the characteristics between the node and neighbor nodes and outputs a weight for neighbor nodes to the GAT network unit.

[0018] Additionally, the dead space policy decision unit according to the above embodiment is characterized by comprising: a policy network unit that receives the occupancy map, design object information, GAT-based embedding, and entropy coefficient as input, calculates a behavioral probability distribution for placement location, direction, or rotation using a policy network, and predicts dead space that may occur after placement of the design object through adjustment of the entropy coefficient and probabilistic or exploratory-based sampling based on the calculated behavioral probability distribution to select a placement policy for a final placement candidate; a reward analysis unit that calculates a reward value through a predefined reward function for the size of the dead space and connection distance after placement of the design object, and reflects the calculated reward value to the policy network; and a feedback generation unit that, after executing placement according to the placement policy, updates the occupancy information and spatial state to reflect them in the occupancy map, and visualizes and outputs the placement result.

[0019] In addition, the entropy control unit according to the above embodiment is characterized by including: an entropy monitoring unit that monitors the reward variance and policy change trends for the deployment policy of the policy network; and an entropy coefficient adjustment unit that tracks the history of recent action choices based on the deployment policy of the dead space policy decision unit to analyze whether a repeating pattern or bias occurs, and determines whether learning is terminated or continued based on whether stable convergence occurs, thereby adjusting the entropy coefficient of the policy network to regulate the balance between exploration and utilization.

[0020] In addition, one embodiment of the present invention is a dead space minimization placement method comprising: a) a step in which, when an occupancy map analysis unit receives information related to a design object, an occupancy map is generated by displaying the placement area occupied by the design object based on the current placement state in a two-dimensional placement area; b) a step in which a GAT analysis unit generates an embedding reflecting the importance between adjacent nodes through a GAT (Graph Attention Networks) based on the connection relationship between the design objects; c) a step in which a dead space policy decision unit receives the occupancy map, design object information, and entropy coefficient as inputs, executes a policy network to predict the dead space that may occur after the placement of the design object, and outputs a placement policy, wherein the embedding reflecting the importance between adjacent nodes through a GAT (Graph Attention Networks) based on the connection relationship between the design objects is reflected in the placement policy; and d) a step in which the dead space policy decision unit executes the placement of the design object according to the placement policy through the placement unit, and after the placement execution, updates the occupancy information and spatial state to reflect them in the occupancy map, and visualizes the placement result to provide feedback;

[0021] In addition, the placement method according to the above embodiment is characterized by further including the step of e) after the placement of the design object, the dead space policy decision unit calculates a compensation value for the size of the dead space and the connection distance through a predefined compensation function, and updates the policy network by reflecting the calculated compensation value.

[0022] In addition, the deployment method according to the above embodiment is characterized by further including the step of f) an entropy control unit monitoring the reward variance and policy change trends for the deployment policy of the policy network, and adjusting an entropy coefficient that controls the balance of exploratoryness and convergence in the action selection process of the policy network based on the entropy and performance indicators of the policy network. Effects of the invention

[0023] The present invention has the advantage of improving space utilization efficiency by performing occupancy map generation, dead space prediction, connection importance analysis, entropy control, and feedback provision through a policy neural network based on a layout in which target objects are placed. Brief explanation of the drawing

[0024] FIG. 1 is an example diagram illustrating the placement results of a placement system according to the prior art. FIG. 2 is a block diagram illustrating a dead space minimization layout system according to an embodiment of the present invention. FIG. 3 is a block diagram illustrating the configuration of the occupancy map analysis unit of the dead space minimization layout system according to the embodiment of FIG. 2. FIG. 4 is a block diagram illustrating the configuration of the GAT analysis unit of the dead space minimization layout system according to the embodiment of FIG. 2. FIG. 5 is a block diagram illustrating the configuration of the dead space policy decision unit of the dead space minimization placement system according to the embodiment of FIG. 2. FIG. 6 is a block diagram illustrating the configuration of the entropy control unit of the dead space minimization layout system according to the embodiment of FIG. 2. FIG. 7 is an example diagram illustrating the occupancy map of a dead space minimization layout system according to the embodiment of FIG. 2. FIG. 8 is a flowchart illustrating a dead space minimization layout method according to one embodiment of the present invention. Specific details for implementing the invention

[0025] Hereinafter, the present invention will be described in detail with reference to preferred embodiments of the invention and the accompanying drawings, under the premise that identical reference numerals in the drawings refer to identical components.

[0026] Before describing specific details for the implementation of the present invention, it should be noted that configurations not directly related to the technical essence of the present invention have been omitted to the extent that they do not detract from the technical essence of the present invention.

[0027] Furthermore, terms or words used in this specification and claims should be interpreted in a meaning and concept consistent with the technical spirit of the invention, based on the principle that the inventor may define the concept of appropriate terms to best describe their invention.

[0028] In this specification, the expression that a part "includes" a certain component means that it does not exclude other components but may include additional components.

[0029] In addition, terms such as "...part," "...unit," and "...module" refer to a unit that processes at least one function or operation, and this can be classified as hardware, software, or a combination of both.

[0030] In addition, the term "at least one" is defined as a term including both singular and plural forms, and it is self-evident that even if the term "at least one" does not exist, each component may exist in the singular or plural form and may mean the singular or plural.

[0031] Hereinafter, a preferred embodiment of a dead space minimization layout system and method according to an embodiment of the present invention will be described in detail with reference to the attached drawings.

[0032] FIG. 2 is a block diagram illustrating a dead space minimization layout system according to an embodiment of the present invention; FIG. 3 is a block diagram illustrating the configuration of an occupancy map analysis unit of a dead space minimization layout system according to an embodiment of FIG. 2; FIG. 4 is a block diagram illustrating the configuration of a GAT analysis unit of a dead space minimization layout system according to an embodiment of FIG. 2; FIG. 5 is a block diagram illustrating the configuration of a dead space policy decision unit of a dead space minimization layout system according to an embodiment of FIG. 2; and FIG. 6 is a block diagram illustrating the configuration of an entropy control unit of a dead space minimization layout system according to an embodiment of FIG. 2.

[0033] As shown in FIGS. 2 to 6, a dead space minimization layout system (100) according to one embodiment of the present invention can generate an occupancy map by displaying the layout area (200, see FIG. 7) occupied by the design object based on the current layout state in a two-dimensional layout area (200) when design object-related information is input.

[0034] Here, the design object may include various electronic and optical elements such as macrocells, standard cells, ports, FPGA modules, process cells, stations, packages, device blocks, components, PCB components, and chiplets of 2.5D packages, as the design object is a physical placement target.

[0035] These design objects are placed in a two-dimensional space and can be used as various physical and logical units that constitute a system.

[0036] Additionally, the placement system (100) can receive an occupancy map through a policy network, predict the dead space that may occur after the placement of the design object, and output a placement policy.

[0037] Additionally, the placement system (100) can generate an embedding that reflects the importance between adjacent nodes through a Graph Attention Network (GAT) based on the connection relationship between design objects, and have it reflected in the placement policy.

[0038] Additionally, the placement system (100) can execute the placement of design objects according to the placement policy and reflect the occupancy status according to the placement execution in the occupancy map.

[0039] Additionally, the deployment system (100) can adjust an entropy coefficient that controls the balance of exploratory and convergent behavior selection processes of the policy network based on the entropy and performance indicators of the policy network.

[0040] Additionally, the placement system (100) can generate feedback information based on the occupancy status and space efficiency information after the placement of the design object and reflect it in the learning process of the policy network.

[0041] To this end, the placement system (100) may be configured to include an occupancy map analysis unit (110), a GAT analysis unit (120), a dead space policy decision unit (130), an entropy control unit (140), and a placement unit (150).

[0042] The occupancy map analysis unit (110) is configured to generate an occupancy map by displaying the placement area (200) occupied by the design object based on the current placement state in a two-dimensional placement area (200) when design object-related information is input, and may be configured to include an input interface unit (111), an occupancy map generation unit (112), and an occupancy information visualization unit (113).

[0043] The input interface unit (111) can generate an array of design object attributes and a list of design objects to be placed based on design object information including unique information, size, and connection target of the design object input by the user.

[0044] In addition, the input interface section (111) can generate a graph data structure in which the design objects are configured as nodes and the connections as edges using the connection relationships between the design objects.

[0045] Additionally, the input interface unit (111) can initialize the layout size and resolution of the placement area (200), that is, make the entire placement area (200) unplaced, so that, for example, the occupancy map is filled with '0'.

[0046] The occupancy map generation unit (112) can display the occupancy area occupied by the design object in the two-dimensional placement area (200) based on the location and size information of the currently placed design object.

[0047] The occupancy map generation unit (112) applies, for example, '1' or a mask to the grid cells occupied by each design object so that occupied areas and unoccupied areas (210, 220, 230) can be displayed as in FIG. 7.

[0048] Additionally, the occupancy map generation unit (112) can generate an occupancy map by updating the occupancy value of the corresponding placement area (200) according to the location and orientation of the newly placed design object.

[0049] The occupancy map updated in the occupancy map generation unit (112) can be used for predicting the next action.

[0050] The occupancy information visualization unit (113) converts the occupancy map generated from the occupancy map generation unit (112) into visualization information such as color and heat distribution and outputs it so that the user can check it.

[0051] The GAT analysis unit (120) is configured to generate embeddings that reflect the importance between adjacent nodes through GAT (Graph Attention networks) based on the connection relationships between design objects and to reflect them in the placement policy, and may be configured to include a GAT network unit (121), a node embedding unit (122), and a weight evaluation unit (123).

[0052] The GAT network section (121) can perform node and neighbor node embedding based on a graph structure composed of nodes and edges, and can perform node embedding by reflecting weights according to importance with neighbor nodes.

[0053] That is, when the GAT network unit (121) receives a connection graph structure and a node characteristic vector, it can perform an attention operation based on the importance of neighboring nodes and output a weighted node embedding.

[0054] The node embedding unit (122) is configured to generate an embedding vector for each node according to design object attributes including the size, type, and unique information of the design object, and expresses the node embedding for each design object by generating an initial embedding vector using the design object attributes.

[0055] The weight evaluation unit (123) calculates an attention score based on the characteristics between the node and the neighbor node to calculate a weight for the neighbor node, and can output the calculated weight for the neighbor node to the GAT network unit (121).

[0056] That is, the weight evaluation unit (123) calculates an attention score using a list of adjacent nodes and the difference in characteristics between nodes, and provides the importance score calculated for neighboring nodes in the form of a probability distribution so that more focus can be placed on the information of the corresponding node.

[0057] The dead space policy decision unit (130) can receive information related to the occupancy map from the occupancy map analysis unit (110) through a policy network, predict the dead space that may occur after the placement of the design object, and output a placement policy.

[0058] Additionally, the dead space policy decision unit (130) can operate to reflect the occupancy status resulting from the deployment execution in the occupancy map, and may be configured to include a policy network unit (131), a reward analysis unit (132), and a feedback generation unit (133).

[0059] The policy network unit (131) receives the occupancy map, design object information, GAT-based embedding, and entropy coefficient from the occupancy map analysis unit (110), GAT analysis unit (120), and entropy control unit (140), enabling the policy network to calculate the action probability distribution for placement location, direction, or rotation.

[0060] That is, the policy network unit (131) can probabilistically determine where it is most advantageous to place the design object by considering the current occupancy map and design object information of the policy network.

[0061] Additionally, the policy network unit (131) can determine the (x, y) position and direction or rotation settings from the judgment result, and can represent these combinations as a single action, and this action can be executed as an optimal arrangement that minimizes dead space or increases connectivity in a given space.

[0062] Additionally, the policy network unit (131) can select a placement policy for the final placement candidate by predicting the dead space that may occur after the placement of the design object through the adjustment of the entropy coefficient and probabilistic or exploratory-based sampling based on the calculated action probability distribution.

[0063] That is, the policy network section (131) reflects the exploratory nature and probability distribution of the policy and is used in actual placement as a specific spatial placement plan, and the result can lead to reward feedback.

[0064] For example, a single action candidate combined with information such as placement position: (x=12, y=7), direction: right among up / down / left / right directions, and rotation angle: 90 degrees can be finally selected and passed to the placement module.

[0065] After the design object is placed, the reward analysis unit (132) calculates a scalar reward value for the size of the dead space and the connection distance through a predefined reward function, and uses the calculated reward value to reflect the adjustment of parameters so that the policy network can learn more effective behavior.

[0066] In addition, the reward analysis unit (132) can normalize and stabilize the main learning direction of the policy network through auxiliary learning goals.

[0067] In other words, by utilizing the current space usage (occupancy status) and the intermediate feature map within the policy network together, it is possible to calculate an auxiliary loss function configured such that a greater loss is incurred when there is more unnecessarily wasted dead space, and a smaller loss is incurred when there is less.

[0068] Through this, important features that are difficult to directly reward can be included in the stable learning direction, and the calculated auxiliary loss is added to the main policy learning loss function to serve as a constraint on space efficiency along with reward maximization.

[0069] The feedback generation unit (133) can update the occupancy information and space status after executing the placement according to the placement policy so that they are reflected in the occupancy map, and can visualize and output the placement results.

[0070] The entropy control unit (140) is configured to control the balance of exploratory power and convergence in the action selection process of a policy network based on the entropy and performance indicators of the policy network, and may include an entropy monitoring unit (141) and an entropy coefficient control unit (142).

[0071] The entropy monitoring unit (141) can monitor the reward distribution and policy change trends for the deployment policy of the policy network.

[0072] That is, the entropy monitoring unit (141) analyzes the change trends of rewards and policies to determine whether the learning has stabilized to the point where it can no longer be improved, and based on this, automatically determines the end point of the learning.

[0073] For example, the smaller the reward variance value, which is the variability of rewards between recent episodes, the more stable the learning performance can be judged, and if the policy change, which is the difference between the previous policy and the current policy, hardly changes, it can be judged that the learning has already converged to some extent and is stable.

[0074] The entropy coefficient control unit (142) can track the history of recent behavioral choices based on the placement policy of the dead space policy decision unit (130) to analyze whether a repeating pattern or bias occurs.

[0075] Additionally, the entropy coefficient control unit (142) can adjust the entropy coefficient of the policy network to control the balance between exploration and utilization by determining whether to terminate or continue learning based on whether stable convergence has occurred.

[0076] That is, the entropy coefficient control unit (142) detects whether only specific actions are repeatedly selected or whether the probability distribution is biased regarding the recent action selection history, and performs frequency analysis or entropy calculation on the action distribution to increase or decrease the entropy coefficient.

[0077] Additionally, the entropy coefficient control unit (142) can increase the entropy coefficient at the beginning of learning to strengthen exploration, gradually decrease the entropy as learning progresses or as the policy stabilizes to induce convergence, and increase the coefficient when the reward stagnates or the entropy is low so that exploration can be restored.

[0078] This ensures that the diversity of behavioral choices and the stability of policy learning are maintained.

[0079] The placement unit (150) is configured to execute the placement of design objects according to the placement policy, and can be applied to the placement of design objects in the corresponding coordinates and direction according to the placement location and the attributes of the design objects.

[0080] Additionally, the placement unit (150) transmits the latest occupancy status, including occupancy information and spatial status regarding the placement result, to the occupancy map analysis unit (110) so that the occupancy map can be updated.

[0081] The following describes a dead space minimization layout method according to one embodiment of the present invention.

[0082] FIG. 8 is a flowchart illustrating a dead space minimization layout method according to an embodiment of the present invention.

[0083] Referring to FIGS. 1 to 8, when the occupancy map analysis unit (110) receives design object-related information from a user, it can generate an occupancy map by displaying the placement area (200) occupied by the design object based on the current placement state in a two-dimensional placement area (200) (S100).

[0084] In step S100, the occupancy map analysis unit (110) can generate an array of attributes of a design object and a list of design objects to be placed based on design object information including unique information, size, and connection target of a design object input by a user, and can display the occupancy area occupied by the design object in a two-dimensional placement area (200) based on the location and size information of the design object.

[0085] In addition, the layout size and resolution of the placement area (200) can be initialized, and a graph data structure can be created by using the connection relationships between design objects to form design objects as nodes and connections as edges.

[0086] Furthermore, the GAT analysis unit (120) can generate an embedding that reflects the importance between adjacent nodes through GAT (Graph Attention networks) based on the connection relationships between design objects from the graph data structure generated in step S100 (S200).

[0087] In step S200, the GAT analysis unit (120) can output a weighted node embedding by performing an attention operation based on the importance of neighbor nodes from the connection graph structure and the node characteristic vector.

[0088] In addition, at step S200, the GAT analysis unit (120) can represent node embeddings for each design object by generating an initial embedding vector for each node according to design object attributes including the size, type, and unique information of the design object.

[0089] Additionally, in step S200, the GAT analysis unit (120) can calculate an attention score based on the characteristics between the node and the neighbor node and calculate a weight for the neighbor node.

[0090] After performing step S200, the dead space policy decision unit (130) can receive an occupancy map, design object information, and an entropy coefficient as inputs, predict the dead space that may occur after the placement of the design object, and execute a policy network to output a placement policy (S300).

[0091] Additionally, at the S300 stage, the dead space policy decision unit (130) can reflect an embedding that reflects the importance between adjacent nodes through GAT based on the connection relationship between design objects in the placement policy.

[0092] Additionally, at step S300, the dead space policy decision unit (130) can probabilistically determine where it is most advantageous to place the design object by considering the policy network's current occupancy map and design object information.

[0093] Additionally, the dead space policy decision unit (130) can determine the position, orientation, or rotation settings of the design object from the judgment result, and by representing this combination as a single action, an optimal arrangement can be executed to minimize dead space or increase connectivity in a given space.

[0094] Additionally, at step S300, the dead space policy decision unit (130) can select a placement policy for the final placement candidate by predicting the dead space that may occur after the placement of the design object through the adjustment of the entropy coefficient and probabilistic or exploratory-based sampling based on the calculated action probability distribution.

[0095] After performing step S300, the dead space policy decision unit (130) executes the placement of design objects according to the placement policy selected in step S300 through the placement unit (150), and after the placement execution, updates the occupancy information and space status according to the placement result so that they are reflected in the occupancy map, and feeds back the placement result so that it is visualized (S400).

[0096] Additionally, after the placement of the design object of step S400, the dead space policy decision unit (130) calculates a scalar compensation value using a predefined compensation function for the size and connection distance of the dead space, and reflects the calculated compensation value in the policy network to update it (S500).

[0097] This allows the policy network to adjust parameters to learn more effective behaviors.

[0098] Additionally, at step S500, the dead space policy decision unit (130) can normalize and stabilize the main learning direction of the policy network by utilizing the current space usage (occupancy status) and the intermediate feature map inside the policy network together to calculate an auxiliary loss function configured such that the more dead space is wasted unnecessarily, the greater the loss, and the less dead space is wasted, the smaller the loss.

[0099] After performing step S500, the entropy control unit (140) monitors the reward variance and policy change trends for the deployment policy of the policy network, and can adjust the entropy coefficient that controls the balance of exploratory and convergent behavior selection in the policy network based on the entropy and performance indicators of the policy network (S600).

[0100] That is, in the S600 stage, the entropy control unit (140) analyzes the change trends of the reward and policy to determine whether the learning has stabilized to the extent that it can no longer be improved, and based on this, automatically determines the learning end point.

[0101] Additionally, at step S600, the entropy control unit (140) tracks the history of recent action selections based on the placement policy of the dead space policy decision unit (130) to analyze whether a repeating pattern or bias occurs, and determines whether to terminate or continue learning based on whether stable convergence occurs, thereby adjusting the entropy coefficient of the policy network to regulate the balance between exploration and utilization.

[0102] Through this, the entropy control unit (140) detects whether only specific actions are repeatedly selected or whether the probability distribution is biased regarding the recent action selection history, and performs frequency analysis or entropy calculation on the action distribution so that the entropy coefficient increases or decreases.

[0103] In the early stages of learning, the entropy coefficient is increased to strengthen exploration, and as learning progresses or the policy stabilizes, the entropy is gradually decreased to induce convergence. When the reward stagnates or the entropy is low, the coefficient is increased to restore explorability, thereby maintaining the diversity of action choices and the stability of policy learning.

[0104] Therefore, space utilization efficiency can be improved by generating an occupancy map, predicting dead space, analyzing connectivity importance, controlling entropy, and providing feedback through a policy neural network based on the layout in which target objects are placed.

[0105] As described above, although the present invention has been explained with reference to preferred embodiments, those skilled in the art will understand that various modifications and changes can be made to the present invention without departing from the spirit and scope of the invention as described in the following claims.

[0106] In addition, the drawing numbers described in the claims of the present invention are provided for clarity and convenience of explanation only and are not limited thereto, and in the process of describing the embodiments, the thickness of lines or the size of components shown in the drawings may be exaggerated for clarity and convenience of explanation.

[0107] Furthermore, the terms described above are defined in consideration of their functions in the present invention, and since these may vary depending on the intentions or practices of the user or operator, the interpretation of these terms should be based on the content throughout this specification.

[0108] Furthermore, it is obvious that a person skilled in the art to which the present invention pertains can make various modifications including the technical concept according to the present invention from the description of the present invention, even if not explicitly shown or described, and such modifications still fall within the scope of the rights of the present invention.

[0109] In addition, the above embodiments described with reference to the attached drawings are described for the purpose of explaining the present invention, and the scope of the present invention is not limited to these embodiments. Explanation of the symbols

[0110] 100 : Deployment System 110 : Occupancy Map Analysis Unit 111: Input Interface Unit 112: Occupancy Map Generation Unit 113: Occupancy Information Visualization Unit 120: GAT Analysis Unit 121: GAT Network Section 122: Node Embedding Section 123: Weight Evaluation Unit 130: Dead Space Policy Decision Unit 131: Policy Network Department 132: Compensation Analysis Department 133 : Feedback generation unit 140 : Entropy control unit 141: Entropy monitoring unit 142: Entropy coefficient control unit 150 : Layout section 200 : Layout area 210, 220, 230: Unoccupied area

Claims

Claim 1 When information related to a design object is input, an occupancy map is generated by displaying the placement area (200) occupied by the design object based on the current placement state in a two-dimensional placement area (200); the occupancy map is received through a policy network to predict the dead space that may occur after the placement of the design object and output a placement policy; an embedding reflecting the importance between adjacent nodes is generated through a Graph Attention Network (GAT) based on the connection relationship between the design objects; an action probability distribution for placement position, direction, or rotation is calculated by receiving the occupancy map, design object information, and the GAT-based embedding; a combination of the placement position, direction, and rotation is determined as a single action; an auxiliary loss function calculated according to the amount of dead space generated is added to the main policy learning loss function using an intermediate feature map inside the policy network and the occupancy state of the occupancy map, and reflected as a constraint on spatial efficiency; and an entropy coefficient that controls the balance of exploratoryness and convergence in the action selection process of the policy network is adjusted based on the entropy and performance indicators of the policy network, and the reward A dead space minimization batch system characterized by a batch system (100) that monitors the trends of variance and policy change and, when it is determined that learning is stable because the reward variance value is small and the policy change does not change, reduces the entropy coefficient to induce convergence, and, when the reward is stagnant or the entropy is low, increases the entropy coefficient to restore explorability. Claim 2 delete Claim 3 A dead space minimization placement system according to claim 1, wherein the placement system (100) generates feedback information based on occupancy status and space efficiency information after the placement of the design object and reflects it in the learning process of the policy network. Claim 4 In claim 1, the placement system (100) comprises: an occupancy map analysis unit (110) that generates an occupancy map by displaying the placement area (200) occupied by the design object based on the current placement state in a two-dimensional placement area (200) when design object-related information is input; a GAT analysis unit (120) that generates an embedding reflecting the importance between adjacent nodes through a GAT (Graph Attention Networks) based on the connection relationship between the design objects and reflects it in the placement policy; an occupancy map-related information including the occupancy map, design object information, and GAT-based embedding is input through a policy network to calculate an action probability distribution for placement position, direction, or rotation, and determines a combination of the placement position, direction, and rotation as a single action; an auxiliary loss function calculated according to the amount of dead space generated by utilizing the intermediate feature map within the policy network and the occupancy state of the occupancy map is added to the main policy learning loss function and reflected as a constraint on spatial efficiency, and predicts the dead space that may occur after the placement of the design object to determine the placement policy A dead space minimization placement system characterized by comprising: a dead space policy decision unit (130) that operates to output and reflect the occupancy status resulting from the placement execution in the occupancy map; an entropy control unit (140) that controls the balance of exploratory power and convergence in the action selection process of the policy network based on the entropy and performance indicators of the policy network, and monitors the trend of reward variance and policy change to induce convergence by decreasing the entropy coefficient when it is determined that learning is stable because the reward variance value is small and the degree of policy change does not change, and increases the entropy coefficient to restore exploratory power when the reward is stagnant or the entropy is low; and a placement unit (150) that executes the placement of the design object according to the placement policy. Claim 5 In claim 4, the occupancy map analysis unit (110) generates a graph structure composed of nodes and edges for the connection relationship between design objects based on design object information and initializes the placement area (200) input interface unit (111); an occupancy map generation unit (112) that displays the occupancy area occupied by the design object in the two-dimensional placement area (200) based on the position and size information of the currently placed design object and generates an updated occupancy map according to the position and direction of the newly placed design object; and an occupancy information visualization unit (113) that converts the generated occupancy map into arbitrary visualization information and outputs it. The dead space minimization placement system is characterized by comprising: an occupancy map analysis unit (110) that generates the occupancy area occupied by the design object in the two-dimensional placement area (200) based on the position and size information of the currently placed design object; and an occupancy information visualization unit (113) that converts the generated occupancy map into arbitrary visualization information and outputs it. Claim 6 In claim 4, the above GAT analysis unit (120) performs node and neighbor node embedding based on a graph structure composed of nodes and edges, and performs node embedding by reflecting weights according to importance with neighbor nodes; a node embedding unit (122) that generates an embedding vector of each node according to attributes including size, type, and unique information of the design object; and a weight evaluation unit (123) that calculates an attention score according to the characteristics between the node and neighbor nodes and outputs a weight for neighbor nodes to the above GAT network unit (121); characterized in that it is a dead space minimization placement system. Claim 7 In claim 4, the dead space policy decision unit (130) receives the occupancy map, design object information, GAT-based embedding, and entropy coefficient as inputs, and the policy network calculates a behavioral probability distribution for placement location, direction, or rotation, and based on the calculated behavioral probability distribution, predicts the dead space that may occur after placement of the design object through adjustment of the entropy coefficient and probabilistic or exploratory-based sampling, and selects a placement policy for the final placement candidate; a reward analysis unit (132) that calculates a reward value through a predefined reward function for the size of the dead space and the connection distance after placement of the design object, and reflects the calculated reward value to the policy network; and a feedback generation unit (133) that updates the occupancy information and spatial state to be reflected in the occupancy map after executing the placement according to the placement policy, and visualizes and outputs the placement result. Claim 8 In claim 4, the entropy control unit (140) comprises: an entropy monitoring unit (141) that monitors the reward variance and policy change trends for the deployment policy of the policy network; and an entropy coefficient adjustment unit (142) that tracks the history of recent action choices based on the deployment policy of the dead space policy decision unit (130) to analyze whether a repeating pattern or bias occurs, and determines whether learning is terminated or continued based on whether stable convergence occurs, thereby adjusting the entropy coefficient of the policy network to regulate the balance between exploration and utilization. Claim 9 a) When the occupancy map analysis unit (110) receives information related to a design object, it generates an occupancy map by displaying the placement area (200) occupied by the design object based on the current placement state in a two-dimensional placement area (200); b) The GAT analysis unit (120) generates an embedding that reflects the importance between adjacent nodes through a GAT (Graph Attention Networks) based on the connection relationship between the design objects; c) The dead space policy decision unit (130) receives the occupancy map, design object information, and GAT-based embedding through a policy network, calculates an action probability distribution for placement position, direction, or rotation, determines a combination of the placement position, direction, and rotation as a single action, and outputs a placement policy; d) The dead space policy decision unit (130) executes the placement of the design object according to the placement policy through a placement unit (150), and after the placement execution, updates the occupancy information and spatial state to reflect them in the occupancy map, and visualizes the placement result to provide feedback. Step; e) after the dead space policy decision unit (130) places the design object, the dead space policy decision unit (130) reflects the reward value calculated according to the size of the dead space and the connection distance to the policy network, and adds an auxiliary loss function calculated according to the amount of dead space generated by utilizing the intermediate feature map and the occupancy status of the occupancy map within the policy network to the main policy learning loss function and reflects it as a constraint on spatial efficiency; f) the entropy control unit (140) monitors the reward variance and the trend of policy change for the placement policy of the policy network, and if it is determined that the learning is stable because the reward variance is small and the policy change does not change, the entropy coefficient is reduced to induce convergence, and if the reward is stagnant or the entropy is low, the entropy coefficient is increased to restore explorability; a dead space minimization placement method comprising Claim 10 delete Claim 11 delete

Images