VLSI mixing unit collaborative layout method based on evolutionary knowledge and landscape perception

Abstract

The invention discloses a VLSI (Very Large Scale Integration) mixing unit collaborative layout method based on evolutionary knowledge and landscape perception, aiming at the defect that the prior art depends on static Netlist information and cannot perceive and optimize dynamic coupling of landscape. The method comprises the following steps: extracting four-dimensional evolution knowledge characteristics of convergence time, an accumulation path, symbol flipping times and direction consistency of a module through a preview stage, constructing a landscape perception type component neighborhood based on clustering analysis, and realizing normal form transformation from structure correlation degree to behavior correlation degree. The method comprises the specific steps that S1, knowledge-driven component neighborhood construction and rehearsal population generation are carried out, and dynamic grouping is generated through initial population legalization, rehearsal behavior data accumulation, feature extraction and clustering; s2, two-stage collaborative optimization of knowledge solidification: firstly optimizing a macro cell group and freezing coordinates, and then optimizing a standard cell group and introducing hot area search; and S3, performing mixed overlapping typing detection and cost evaluation, calculating three types of overlapping areas through hierarchical detection, and constructing a total cost function. According to the method, the implicit coupling relation is effectively captured, the grouping accuracy and optimization efficiency are improved, the zero-overlapping and low-HPWL targets are achieved, and the method is suitable for a high-dimensional noise-containing VLSI layout scene.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of very large scale integrated circuit (VLSI) physical design, and particularly relates to a high-efficiency optimization method based on cooperative evolution for macro cell and standard cell collaborative layout. BACKGROUND

[0002] In VLSI physical design, layout optimization needs to find the optimal coordinate arrangement for macro cells (large area, small quantity) and standard cells (large quantity, small size) in the chip plane to minimize half-perimeter wire length (HPWL) and eliminate physical overlap. The cooperative evolution (Cooperative Coevolution, CC) algorithm decomposes the high-dimensional problem into multiple low-dimensional sub-problems through the "divide and conquer" idea, and its performance is highly dependent on the division quality of the component neighborhood. The existing technology usually constructs groups based on the static structural information (such as modularity, pairing weight) of Netlist, but cannot perceive the implicit coupling relationship emerging in the optimization process, for example, the dynamic association formed by multiple modules cooperatively bypassing the same macro cell obstacle. This static grouping strategy is easily dominated by the large penalty item of macro cell overlap in noisy, high-dimensional layout, leading to degradation of grouping quality and optimization into local optimum.

[0003] The existing mainstream methods mainly include simulated annealing heuristic algorithm, single population differential evolution or genetic algorithm, and gradient or force-oriented method. However, these technologies expose the following core deficiencies in actual engineering applications: first, the initial solution quality is low, and the random initialization strategy leads to a large amount of evaluation cost wasted in the early stage on poor solution space far from the optimal solution region, which reduces the overall search efficiency; second, high-dimensional coupling is difficult to solve, the complex coupling relationship between macro cells and macro cells, macro cells and standard cells, and standard cells and standard cells makes the full-dimensional disturbance easily destroy the formed high-quality substructure, and fall into local optimum; third, the processing cost of overlap is high, and the overlap detection and repair mechanism based on complete iteration presents square computational complexity in large-scale scenarios, which becomes a performance bottleneck; in addition, the function of the operator is single, and the same evolutionary operator needs to undertake the dual tasks of global exploration and local convergence, which easily leads to premature convergence of the population or solution quality shock; finally, there is a lack of type strategy, macro cells and standard cells have order of magnitude difference in geometric scale, connection weight and arrangement constraint, but are treated uniformly, the search mechanism is not designed specifically, which leads to unbalanced resource allocation and low optimization efficiency. More importantly, the existing collaborative layout method relies heavily on static Netlist information (such as module degree and pair weight) when constructing component neighborhood, and cannot reflect the implicit coupling emerging in the optimization landscape. For example, two standard cells with weak static connection may show strong behavioral correlation because they must bypass the same large macro cell, and the traditional static grouping cannot identify such "evolutionary behavioral correlation degree", leading to low grouping quality and optimization into local optimum. Therefore, the industry urgently needs an advanced grouping mechanism that can perceive dynamic fitness landscape and accurately identify the real interaction between variables to break through the optimization bottleneck in high-dimensional and noisy layout scenarios. SUMMARY

[0004] The purpose of the present application is to overcome the defects of the existing VLSI collaborative layout method that relies on static Netlist and cannot perceive the optimization landscape, and to provide a VLSI mixed cell collaborative layout method based on evolutionary knowledge and landscape perception. In the pre-evolution stage, the method actively explores the convergence behavior of the modules through differential evolution (DE), extracts four-dimensional evolutionary knowledge features (convergence time , cumulative path , number of sign flips , direction consistency ), and constructs a landscape perception type component neighborhood based on cluster analysis for subsequent two-stage type collaborative optimization. The core innovation of the present application is to replace the traditional static grouping with a dynamic behavior perception type grouping, which first realizes the paradigm shift from "structural correlation degree" to "behavioral correlation degree". Specifically, the method includes the following steps:

[0005] S1: Knowledge-driven component neighborhood construction and pre-evolution population generation

[0006] In this stage, the initial population is legalized, behavior data is accumulated through pre-evolution, evolution knowledge features are extracted, and a landscape perception component neighborhood is generated through clustering, providing a dynamic grouping basis for subsequent optimization. The core process is shown in detail in FIG. 1, and each sub-step forms a closed loop of “generation-evolution-extraction-clustering”.

[0007] S1.1 Initial population generation and boundary clipping

[0008] In the chip plane, pseudo-random number generators are used to generate initial coordinates for all macro cells and standard cells; a function is called to perform boundary clipping to ensure that each module is completely located within the chip plane, and the corresponding population size is set.

[0009] S1.2 Pre-evolution and mean trajectory accumulation

[0010] A differential evolution algorithm is started to perform pre-evolution, with a pre-evolution budget of 20% of the total evaluation budget, and parameters SSP=0.90 and CBV_SVAL=8. Every 8 generations, the optimal solution of the current population is executed for macro cell fast legalization, and the mean of the repaired population is added to the “mean trajectory”. This operation can weaken the interference of the macro cell overlap large penalty on subsequent behavior feature extraction.

[0011] S1.3 Evolution knowledge feature extraction

[0012] Four-dimensional evolution knowledge features are extracted for each coordinate dimension from the mean trajectory, and the extraction logic is shown in detail in Figure 2 , which includes:

[0013] S1.3.1 Convergence time Feature extraction: take the median of the mean value changes in the first 1 / 3 of the pre-evolution generations, and determine the convergence generation number for each dimension through adaptive double threshold , , combined with a sliding window;

[0014] S1.3.2 Cumulative path Feature extraction: absolute accumulation of historical displacement for each coordinate dimension;

[0015] S1.3.3 Number of sign flips Feature extraction: count the frequency of sign changes of the velocity vector for each dimension;

[0016] S1.3.4 Direction consistency Feature extraction: calculate the cosine similarity between the cumulative displacement vector of each dimension and the global average direction, mapped to [0,1]; finally, the four-dimensional features are normalized by min-max and concatenated into a feature vector.

[0017] S1.4 Landscape-aware component neighborhood generation

[0018] K-means clustering is performed on the normalized four-dimensional feature vectors; after clustering, "balanced slicing" is performed to divide the target group according to the target group size: the target group size of macro units and the target group size of standard units are specified to avoid "oversized clusters diluting the budget" and "soft constraints pulling back", and strong edge pairs are preferentially adjusted to the same cluster to avoid "mis-separation of strong edge modules"; finally, the neighborhood of landscape-aware components is generated, and elite individuals in the external archives are saved as the initial high-quality population for subsequent optimization.

[0019] S2: Two-stage collaborative optimization of knowledge solidification

[0020] S2.1 Macrocell Priority Optimization: Based on the component neighborhood generated in S1, the enhanced SaNSDE algorithm is used for the first stage of optimization: priority optimization of macrocell groups is given, and budget adaptive allocation is adopted (weighted by group size and internal edge weight).

[0021] S2.2 Macrocell Freezing: After the first stage, perform high-intensity vector repulsion force iterative repair on the macrocells (iteration count ≥ 100, tolerance ≤ 1e-9) to ensure that the macro-macro overlap area is reduced to below zero; then freeze the macrocell coordinates as global constants.

[0022] S2.3 Standard Element Optimization and Hot Zone Search: Perform the second stage of optimization: Perform SaNSDE perturbation optimization on the standard element group; at the same time, introduce the "hot zone" local search mechanism - randomly select a subgroup with high edge density, construct a weighted difference vector (the weight is 1 in the hot zone and 0 elsewhere), generate experimental solutions and evaluate them, and accept them if the cost is reduced.

[0023] S3: Hybrid Overlapping Classification Detection and Cost Assessment

[0024] The three types of overlap areas are accurately calculated by using a layered overlap detection strategy. The total cost function is constructed by combining HPWL and the optimization termination condition is clearly defined.

[0025] S3.1 Hybrid Overlap Detection: When the total number of modules is ≤450, the overlap area is calculated using full matrix broadcasting; otherwise, regular grid bucket partitioning combined with nine-square neighborhood enumeration is used to accumulate the overlap areas of macro-macro, standard-standard, and hybrid types respectively.

[0026] S3.2 Cost Assessment and Termination Judgment: The total cost is defined as the sum of HPWL and the three types of overlapping penalties, with the macro-macro penalty weight set to and the standard and hybrid penalty weight set to . The termination conditions are: reaching the maximum number of assessments, the macro cell overlap area being less than the set threshold, or the cost improvement rate being less than 0.01% after multiple consecutive assessments. Attached Figure Description

[0027] Figure 1 Build a detailed flowchart for the knowledge-driven component neighborhood in Phase S1;

[0028] Figure 2 A schematic diagram of the evolutionary knowledge feature extraction process.

[0029] Figure 3 This is a flowchart illustrating the overall process of this invention. Detailed Implementation

[0030] Within a 300×300 chip plane, 5 macrocells are set, each with a size of 45×45; 100 standard cells are set, each with a size of 2×2. The netlist has an average of 3 edges per module. The optimization objective is to minimize the half-perimeter length (HPWL) and achieve zero overlap, with a maximum of 10,000 evaluations. The invention process logic is detailed in Figure 1, the overall flowchart.

[0031] S1.1 Initial Population Generation

[0032] Using a pseudo-random number generator with a seed of 12345, initial coordinates are generated for all modules within the range [0, 300] × [0, 300]. All modules include macrocells and standard cells. Boundary clamping is performed to ensure that each module is completely within the chip plane, with no coordinates exceeding the boundaries. The population size is set to NP = 60.

[0033] S1.2 Pre-evolution and Knowledge Accumulation

[0034] The pre-evolution algorithm is initiated, with a pre-evolution budget of 20% of the total evaluation budget. The total evaluation budget is 10,000 iterations, therefore the pre-evolution budget is 2,000 evaluations. Parameters SSP=0.90 and CBV_SVAL=8 are set. Every 8 generations, a macrocell fast legalization operation is performed on the current population's optimal solution, and the legalized and repaired population mean is appended to the historical mean record (mean trajectory). This mechanism effectively avoids interference from the large penalty term for macrocell overlap on subsequently extracted evolutionary knowledge features. The weight of the large penalty term for macrocell overlap is 1e7, and the evolutionary knowledge features include convergence time. Cumulative path This ensures the purity of the feature data.

[0035] S1.3 Evolutionary Knowledge Feature Extraction

[0036] Extracting four-dimensional evolutionary knowledge features from historical mean records for each coordinate dimension; for details of the extraction logic, please refer to [link to relevant documentation]. Figure 3 The steps are as follows: convergence time Take the mean change of the first 1 / 3 of the pre-evolutionary generations. The first 1 / 3 of the pre-evolutionary generations is approximately 83 generations. The mean change is expressed as... This indicates that the median of the change in mean is calculated. An adaptive dual threshold is then set based on this median, where... , The convergence algebra for each dimension is determined using a sliding window, with the parameters of the sliding window being... The convergent algebra is the convergence time. .

[0037] Cumulative path The historical displacement of each coordinate dimension is absolutely accumulated, and the final value range is [0, 1200].

[0038] Number of symbol flips : Calculate the sign change frequency of the velocity vector in each coordinate dimension, with a value range of [0, 180].

[0039] Consistency of direction Calculate the cosine similarity between the cumulative displacement vector of each coordinate dimension and the global average displacement direction, and map the result to the interval [0,1].

[0040] The above four-dimensional features are normalized using min-max normalization. These four features are the convergence time, etc. Cumulative path Number of symbol flips Consistency of direction After processing, they are concatenated into a 4-dimensional feature vector.

[0041] S1.4 Landscape-Perception Grouping

[0042] 1. Set target group size: The target group size for macrocells is 2, and the target group size for standard cells is 10. k-means clustering: Perform k-means clustering on the 4-dimensional feature vectors of macrocells and standard cells respectively.

[0043] 2. k-means clustering. The number of macrocell clusters. This value is obtained by rounding up to accommodate the grouping requirement of 5 macrounits; standard unit cluster number To accommodate the grouping requirements of 100 standard units.

[0044] 3. Balanced slicing: The large clusters obtained from clustering are divided into multiple subgroups according to the target group size to avoid the problem of "oversized clusters diluting the optimization budget" and ensure that the size of each subgroup is approximately equal.

[0045] 4. Soft Constraint Pullback: Calculate PAIR_W, the edge weight, for all modules and determine its 97.5th percentile threshold, which is approximately 1.8. For strong edge pairs with PAIR_W ≥ this threshold, perform lightweight grouping adjustment, prioritizing their placement in the same subgroup to ensure the integrity of critical edge relationships.

[0046] The final grouping results are as follows: there are 3 macrounit groups with a module distribution of 2+2+1; and 10 standard unit groups with an average of 10 modules per group. Simultaneously, the top 10 elite individuals from the external archives during the pre-evolution stage are saved as the initial high-quality population for subsequent optimization.

[0047] S2.1 Macrocell Priority Optimization

[0048] Allocate 45% of the total evaluation budget, which is 10,000 evaluations. Therefore, the budget for this phase is 4,500 evaluations, used for macrocell optimization.

[0049] 1. Optimized Algorithm: For each macrocell, perform the enhanced SaNSDE algorithm in groups, with the algorithm parameters set to the scaling factor. Crossover probability .

[0050] 2. Budget Allocation: Allocate the optimization budget for each macrocell group according to the following formula: Where base is the basic budget coefficient. This parameter represents the group size weight; This represents the number of modules within a group. This parameter is the edge weight coefficient; The average PAIR_W of modules within a group.

[0051] 3. Legality Guarantee: After each macrocell group optimization is completed, 15 macrocell repair operations are immediately performed to ensure that there is no significant overlap between macrocells and to maintain the legality of the optimization process.

[0052] S2.2 Macrocell Freeze

[0053] After the first stage (macrocell optimization) is completed, high-intensity vector repulsion force iterative repair is performed on the macrocells. The number of iterations is set to 120, the repair step size to 0.35, and the convergence tolerance to ≤ 1e-9, until the macro-overlap area is reduced to < 1e-10. The repaired macrocell coordinates are then saved as... and ,in It is the fixed x-axis coordinate of the macro element. The y-axis is the fixed coordinate of the macro-element, and both serve as global immutable constraints in the second stage (standard element optimization).

[0054] S2.3 Standard Element Optimization and Hotspot Search

[0055] The remaining 35% of the total evaluation budget is allocated, with a total evaluation budget of 10,000 evaluations. Therefore, the budget for this phase is 3,500 evaluations, used for standard cell group optimization.

[0056] Basic optimization: The optimization logic for standard cell grouping is consistent with S2.1, both employing the enhanced SaNSDE algorithm, with only the budget allocation formula being adjusted. This is to adapt to the edge weight characteristics of standard cells. Hotspot search: A local search for "hotspots" is triggered every 100 evaluations, with the following steps:

[0057] ① Randomly select a standard unit group as the hot zone. ;

[0058] ② Randomly select 3 individuals from the current population, denoted as 3. , , ;

[0059] ③ Construct the test vector: ,in The best individual in the current population is represented by w, which is the weight vector, and the hot zone is represented by w. The weight of the corresponding dimension is 1, and the weight of the other dimensions is 0;

[0060] ④ To After implementing boundary clamping, evaluate its total cost; if the total cost is lower than the current optimal value, accept the trial vector and update the optimal individual in the population.

[0061] S3.1 Fractal Overlap Detection

[0062] Select an overlap detection strategy based on the total number of modules: Current total number of modules It consists of 5 macrocells and 100 standard cells, and this value is ≤ 450. Therefore, the full matrix broadcasting method is used to calculate the overlap area; if The calculation is performed using the "regular grid bucket partitioning + nine-square neighborhood enumeration" method. The overlapping areas of the three types are accumulated separately. , and ,in For macro-macro overlap area, For standard-standard overlap area, For macro-standard mixed overlap area.

[0063] S3.2 Cost Assessment and Termination

[0064] Cost function: The total cost consists of HPWL and three types of overlapping penalties, as shown in the following formula:

[0065]

[0066] The macro overlap penalty has a weight of 1e7 to prioritize eliminating macro cell overlap; the standard-standard and macro-standard mixed overlap penalty has a weight of 1e3 to adapt to the size characteristics of standard cells. Termination condition: Optimization terminates when any of the following conditions are met:

[0067] ① The number of assessments has reached the maximum budget, which is 10,000.

[0068] ② Macrocell overlap area ;

[0069] ③ If the cost improvement rate after 1000 consecutive evaluations is less than 0.01%, the optimization tends to converge.

[0070] III. Implementation Results

[0071] Compared with existing technologies, this invention demonstrates significant advantages in VLSI hybrid cell cooperative layout optimization. Firstly, this invention introduces convergence time... Cumulative path Number of symbol flips and directional consistency The four-dimensional evolutionary knowledge features enable precise perception of the dynamic behavioral relationships of modules during the optimization process, effectively capturing implicit coupling relationships that the Netlist structure cannot reflect. Compared with traditional grouping methods that rely on static edge weights, it has stronger adaptability and grouping accuracy. Secondly, by adopting a "legitimization before learning" mechanism combined with adaptive dual-threshold judgment, it can actively isolate the interference of large penalties for macrocell overlap in the pre-evolution stage, ensuring the purity and robustness of behavioral feature extraction and significantly improving the stability of the algorithm in noisy high-dimensional scenarios. In addition, through the balanced slicing after clustering and the soft constraint pull-back strategy for strong edges, this invention ensures the uniformity of the component neighborhood size and the integrity of key coupling relationships, avoiding optimization efficiency losses caused by "oversized clusters diluting the budget" or "mis-separation of strongly connected modules". Finally, at the execution level, high-intensity macrocell repair and coordinate freezing, standard cell hot zone weighted perturbation search, and hierarchical overlap detection mechanism together constitute an efficient closed-loop optimization process, which not only has low computational overhead and fast convergence speed, but also achieves the dual goals of zero overlap and low HPWL within a limited evaluation budget, with overall performance superior to existing static grouping collaborative optimization schemes.

Claims

1. A VLSI hybrid unit cooperative layout method based on evolutionary knowledge and landscape perception, characterized in that, Includes the following steps: S1 Knowledge-driven component neighborhood construction and pre-evolutionary population generation: Through initial population legitimization, pre-evolutionary accumulation of behavioral data, extraction of evolutionary knowledge features, and clustering to generate landscape-aware component neighborhoods, a dynamic grouping basis is provided for subsequent optimization; S2 Knowledge Solidification Two-Stage Collaborative Optimization: Employing a phased strategy of macrocell priority optimization, macrocell freezing, standard cell optimization, and hot zone search; S3 Hybrid Overlap Classification Detection and Cost Assessment: Accurately calculates the overlap area of ​​three types of overlap using a layered overlap detection strategy, constructs a total cost function based on HPWL, and clarifies the optimization termination condition. According to claim 1, the VLSI hybrid unit cooperative layout method based on evolutionary knowledge and landscape perception is characterized in that the S1 knowledge-driven component neighborhood construction and pre-evolutionary population generation include: S1.1 Initial Population Generation and Boundary Clamping: Generate initial coordinates for all macrocells and standard cells within the chip plane, and perform boundary clamping to ensure that each module is completely located within the chip plane; S1.2 Pre-evolution and accumulation of mean trajectory: Start the differential evolution algorithm to perform pre-evolution, and perform macro-unit fast legalization on the current population's optimal solution every set number of generations, and add the repaired population mean to the mean trajectory; S1.3 Evolutionary knowledge feature extraction: Extract four-dimensional evolutionary knowledge features from the mean trajectory, including convergence time, cumulative path, number of sign flips, and direction consistency. S1.4 Generation of Landscape-Aware Component Neighborhood: Perform k-means clustering on the normalized four-dimensional feature vector, and perform balanced slicing and soft constraint pull-back to generate the neighborhood of landscape-aware components.

2. The VLSI hybrid unit cooperative layout method based on evolutionary knowledge and landscape perception according to claim 1, characterized in that, In S1.2, the pre-evolution budget is 20% of the total evaluation budget. The parameters SSP=0.90 and CBV_SVAL=8 are set. The macrocell fast legalization operation is performed every 8 generations to avoid interference from large penalty items due to macrocell overlap.

3. The VLSI hybrid unit cooperative layout method based on evolutionary knowledge and landscape perception according to claim 1, characterized in that, The S1.3 evolutionary knowledge feature extraction includes: Convergence time feature extraction: taking the median of the mean change of the first 1 / 3 of the pre-evolutionary generations, and determining it through adaptive dual thresholds, the calculation formula is as follows: , The convergence algebra of each dimension is determined by combining a sliding window; cumulative path feature extraction: the historical displacement of each coordinate dimension is absolutely accumulated; sign flip number feature extraction: the sign change frequency of the velocity vector of each dimension is statistically analyzed; direction consistency feature extraction: the cosine similarity between the cumulative displacement vector of each dimension and the global average direction is calculated and mapped to [0,1]; finally, the four-dimensional features are concatenated into a feature vector after min-max normalization.

4. The VLSI hybrid unit cooperative layout method based on evolutionary knowledge and landscape perception according to claim 1, characterized in that, In S1.4, the target group size of macrocells is specified as 2, and the target group size of standard cells is specified as 10. After clustering, balanced slicing is performed to divide the cells according to the target group size. Soft constraints are used to calculate PAIR_W between all modules, determine its 97.5% quantile threshold, and prioritize the adjustment of strong connection pairs with PAIR_W ≥ the threshold to the same cluster.

5. The VLSI hybrid unit cooperative layout method based on evolutionary knowledge and landscape perception according to claim 1, characterized in that, The two-stage collaborative optimization of S2 knowledge solidification includes: S2.1 Macrocell Priority Optimization: Based on the component neighborhood generated in S1, the enhanced SaNSDE algorithm is used for the first stage of optimization, and the budget is allocated by weighted average according to the group size and the weight of the internal connection. S2.2 Macrocell Freezing: After the first stage, perform high-intensity vector repulsion force iterative repair on the macrocells to ensure that the macro-macro overlap area is reduced to below zero, and freeze the macrocell coordinates as global constants; S2.3 Standard Element Optimization and Hot Zone Search: Perform the second stage of optimization, apply SaNSDE perturbation optimization to the standard element group, and introduce a local hot zone search mechanism.

6. The VLSI hybrid unit cooperative layout method based on evolutionary knowledge and landscape perception according to claim 1, characterized in that, The optimized budget allocation formula for each macrocell group in S2.1 is as follows: Where base is the basic budget coefficient. This parameter represents the group size weight; This represents the number of modules within a group. This parameter is the edge weight coefficient; The average PAIR_W of modules within a group.

7. The VLSI hybrid unit cooperative layout method based on evolutionary knowledge and landscape perception according to claim 1, characterized in that, In S2.2, the number of iterations is set to ≥100, the convergence tolerance is set to ≤1e-9, until the macro overlap area is <1e-10; the repaired macro cell coordinates are saved as a globally immutable constraint.

8. The VLSI hybrid unit cooperative layout method based on evolutionary knowledge and landscape perception according to claim 1, characterized in that, In S2.3, the hot zone search is triggered once after a set number of evaluations are completed, and a standard unit group is randomly selected as the hot zone to construct the test vector.

9. The VLSI hybrid unit cooperative layout method based on evolutionary knowledge and landscape perception according to claim 1, characterized in that, The S3 hybrid overlap typing detection and cost assessment includes: S3.1 Hybrid Overlap Detection: When the total number of modules is ≤450, the overlap area is calculated using full matrix broadcasting; otherwise, regular grid bucket partitioning combined with nine-square neighborhood enumeration is used to accumulate the overlap areas of macro-macro, standard-standard, and hybrid types respectively. S3.2 Cost Assessment and Termination Determination: The total cost is defined as the sum of HPWL and the three types of overlapping penalties. The termination conditions are reaching the maximum number of assessments, the macrocell overlap area being less than the set threshold, or the cost improvement rate being less than 0.01% in multiple consecutive assessments.

Images