A multi-stage, evolutionary stacking-based system for accurate and agile effort estimation.

Abstract

A system for effort estimation in agile software development using multi-stage evolutionary stacking, consisting of: a data acquisition module configured to retrieve software effort records from one or more data set repositories containing historical data from software development projects with characteristics and actual effort values; a data preprocessing module that is operationally connected to the data acquisition module and is configured to receive the aforementioned software effort data sets from the data acquisition module, cleans the received data by removing inconsistencies with missing target values, and normalizes numerical input characteristics to a common range; a first-level ensemble module connected to the data preprocessing module, wherein the first-level ensemble module comprises a variety of heterogeneous basic learners, including a Random Forest model, a Support Vector Regression model, and an Extreme Gradient Boosting model, which generate predictions from each of the heterogeneous basic learners using the preprocessed data sets received from the data preprocessing module; A genetic algorithm optimization module connected to the first layer's ensemble module, configured to: encode weights as a normalized real-valued vector for each of the heterogeneous base learners; apply a fitness function to minimize the mean squared validation error and derive an optimal weight vector; assign optimized weights to the predictions of each of the heterogeneous base learners; and generate weighted predictions based on the optimized weights. a second-level meta-learning module connected to the optimization module of the genetic algorithm, configured to receive the weighted predictions from the optimization module of the genetic algorithm, processes the weighted predictions using a deep multilayer perceptron neural network to learn complex patterns and nonlinear interactions, and generates a final effort estimate for the software; an output processing module connected to the second-level meta-learning module, configured to receive the final effort estimate for the software and process and visualize the data to improve user understanding; and a user interface connected to the output processing module to receive the processed final effort estimate for the software, wherein the user interface is configured to display the processed and visualized final effort estimate for the software.

Description

AREA OF INVENTION

[0001] The present disclosure relates to a system for effort estimation in agile software development using deep learning, in particular a system for effort estimation in agile software development using multi-stage evolutionary stacking. More precisely, the present invention relates to an intelligent AI system that combines several predictive models, optimizes them using evolutionary methods, and refines the predictions through deep learning in order to accurately estimate the effort of software projects. BACKGROUND OF THE INVENTION

[0002] Agile software development follows an iterative and incremental approach that spans multiple sprints with user stories. Agile projects run in iterative sprint sequences, with each sprint having its own development and deployment cycle. Accurate effort estimation is crucial for completing projects on time and within budget. However, inaccurate effort estimation during the planning phase has proven to be a major cause of project failure, as estimates are made before the necessary experience for precise calculations is available.

[0003] Traditional estimation methods such as expert opinions, user story points, and algorithmic approaches like COCOMO have proven less effective in agile environments due to constant changes and rapid, iterative planning. Although current literature indicates a growing trend toward machine learning for effort estimation in agile software development, existing methods still have significant limitations.

[0004] Early research explored traditional and standalone machine learning models, including support vector regressors, neural networks, and regression-based approaches. However, these models exhibited inconsistencies in their predictions in early stages and were unable to establish nonlinear relationships between datasets, with performance being highly dependent on the dataset's characteristics.

[0005] Ensemble-based methods have shown performance improvements over single models. Studies have confirmed that ensemble methods, particularly Random Forest, are more reliable than deep learning or standalone regression models. However, this work primarily focused on performance improvements without addressing the optimal integration and weighting of different learning algorithms. This limited their adaptability and robustness with heterogeneous and agile datasets.

[0006] Several studies have proposed metaheuristic algorithms to address challenges in hyperparameter and feature selection. These studies have employed optimization algorithms such as the Grey Wolf Optimizer, the Ant Lion Optimizer, and particle swarm optimization. While these approaches have demonstrated that metaheuristics can optimize individual learning models, they have not addressed the central challenge of optimizing the integration of multiple models within an ensemble.

[0007] Multi-stage and stacked architectures have shown great potential by combining the results of different learning algorithms. However, previous studies predominantly use simple linear meta-learning algorithms with manually assigned or grid-searched weights. These ensemble and stacking techniques rely on manually assigned, heuristic, or fixed weighting schemes, leading to inconsistent results across different datasets. The selection and weighting of the base models are usually ad hoc, and an optimization mechanism is lacking to determine how individual predictors should contribute to the final estimates.

[0008] Existing ensemble approaches based on machine learning use static or heuristic weighting combinations. This leads to inconsistent predictions across different datasets because they inadequately capture nonlinear model interactions and cannot dynamically adapt to dataset variations. The high complexity and manual optimization effort of these approaches result in static ensemble configurations that cannot dynamically adapt to different project scenarios.

[0009] It is clear from the preceding discussion that a robust effort estimation system is required. The present invention therefore offers a corresponding approach. It addresses static weight combinations through a GA-driven evolutionary stacking approach that automates and optimizes the weight selection between different base models. The invention introduces an evolutionary optimization mechanism that addresses linear meta-learners with predefined or manually adjusted weights, as used in traditional stacking and ensemble approaches. The proposed mechanism is efficient enough to determine the contribution of each base learner to improving robustness and prediction accuracy across a wide range of project scenarios.

[0010] The present invention provides a multi-layer evolutionary stacking system (MLES) with ensemble modeling that significantly improves performance compared to single models. The MLES-based system utilizes stacked ensemble learning and integrates RF, SVR, and XGBoost as base learners in the first layer. A general analysis (GA), applied to the base learner's output, serves to predict weight through model optimization. A deep multi-layer program (MLP) is used as the second layer and further processes the base learner's output. This multi-layered modeling enables more accurate effort estimations because it combines the strengths of different models and optimization techniques. The goal of the proposed system is to reduce the variance between estimated and actual effort. The results are evaluated using the performance metrics MAE, sMAPE, RMSE, and R. 2-Score for the predictability of effort estimates in agile software projects compared to increase the success rate. SUMMARY OF THE INVENTION

[0011] The present disclosure relates to a system for effort estimation in agile software development using multi-stage evolutionary stacking. The architecture of the proposed system integrates heterogeneous basic learning methods (RF, SVR, XGBoost) using a multi-layered ensemble framework specifically designed for effort estimation in agile software development. The system further includes a GA-based adaptive weight optimization mechanism, wherein the GA is directly embedded in the prediction pipeline and develops real-valued ensemble weights. This invention utilizes a fitness function based on mean squared error (MSE) validation and normalization-driven evolutionary fitting, rather than simply optimizing hyperparameters. The genetic algorithm (GA) generates weighted prediction vectors that serve as inputs for a meta-learner.The system then learns hierarchical, nonlinear patterns from the signals generated by the GA. This makes the GA an integral part of the decision-making process within the prediction function, and not just a preprocessing step.

[0012] This disclosure describes a system for agile effort estimation in software development using multi-stage evolutionary algorithms. The system comprises: a data acquisition module that retrieves software effort datasets from one or more data repositories. These datasets contain historical project data with attributes and actual effort values. The system also includes a data preprocessing module connected to the data acquisition module that receives the datasets, resolves inconsistencies by removing missing target values, and normalizes numerical input attributes to a common range. Finally, the system includes a first-level ensemble module connected to the data preprocessing module. This first-level ensemble module consists of several heterogeneous base models, including a random forest model, a support vector regression model, and an extreme gradient boosting model.Each of these base models generates predictions based on the preprocessed datasets received by the data preprocessing module. The system also includes a genetic algorithm optimization module connected to the first-layer ensemble module. This module is configured to encode the weights for each of the heterogeneous base learners as a normalized real-valued vector, apply a fitness function to minimize the mean squared validation error and derive an optimal weight vector, assign optimized weights to the predictions of each of the heterogeneous base learners, and generate weighted predictions based on these optimized weights. The system also includes a second-layer meta-learning module connected to the genetic algorithm optimization module.This module is configured to receive weighted predictions from the genetic algorithm optimization module, process them using a deep multilayer perceptron network to learn complex patterns and nonlinear interactions, and generate a final effort estimate for the software. The system also includes an output processing module connected to the second-layer meta-learning module. This module is configured to receive the final effort estimate for the software and process and visualize the data to enhance user understanding. A user interface connected to the output processing module receives the processed final effort estimate for the software and is configured to display the processed and visualized final effort estimate.

[0013] The objective of the present invention is to provide an MLES-based system for predicting the effort required for agile software development. In this proposed system, a GA-based evolutionary optimization mechanism is introduced into the prediction pipeline, wherein this mechanism adjusts the base learner weights within a multi-stage stack architecture.

[0014] Another objective of the present invention is to provide a robust effort estimation system for agile Scrum projects, which focus on iterative planning and execution, in order to ensure project success and to establish a strong bond between IT service providers and their customers.

[0015] A further objective of the present invention is to solve the problem of static weight combination by means of a genetic algorithm-driven evolutionary stacking approach that automates and optimizes the weight selection between different base models. The invention aims to integrate different predictive models for software effort estimation by introducing an evolutionary optimization mechanism that determines the contribution of each base learner.

[0016] Another objective of the present invention is to improve the robustness and predictive accuracy of the effort across a wide range of project scenarios through the dynamic development of optimal weighting combinations of heterogeneous basic learners.

[0017] Another objective of the present invention is to capture nonlinear and hierarchical dependencies between predictions by means of a deep multilayer perceptron meta-learner that operates with genetic algorithm-optimized outputs, thereby improving performance compared to single models by using stacked ensemble learning that combines the strengths of different models and optimization techniques.

[0018] However, another objective of the present disclosure is to increase the predictability of effort estimates and the success rate of agile software projects through multi-layered modeling, which provides more accurate effort estimates.

[0019] To further clarify the advantages and features of the present disclosure, the invention is described in more detail with reference to specific embodiments illustrated in the accompanying drawings. It is understood that these drawings merely show typical embodiments of the invention and are therefore not to be understood as limiting its scope of protection. The invention is described and explained in more detail and with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE IMAGES

[0020] These and other features, aspects and advantages of the present disclosure will be better understood when the following detailed description is read with reference to the accompanying drawings, in which identical symbols represent identical parts, wherein: Fig. Figure 1 illustrates a block diagram of a system for effort estimation in agile software development using multi-stage evolutionary stacking according to an embodiment of the present disclosure; and Fig. Figure 2 shows a block diagram of the proposed effort estimation system using multi-stage evolutionary stacking according to an embodiment of the present disclosure.

[0021] Furthermore, those skilled in the art will recognize that the elements in the drawings are simplified and not necessarily drawn to scale. For example, the flowcharts illustrate the process by highlighting the main steps to facilitate understanding of this disclosure. With regard to the construction of the device, one or more components may be represented in the drawings by conventional symbols. The drawings may show only those specific details relevant to understanding the embodiments of this disclosure, so as not to clutter the drawings with details that are already apparent to those skilled in the art from the description contained herein. DETAILED DESCRIPTION:

[0022] To facilitate understanding of the principles of the invention, reference is made below to the embodiment illustrated in the drawings, which is described using specific terms. It is understood, however, that this does not limit the scope of protection of the invention. Rather, modifications and further developments of the illustrated system, as well as further applications of the inventive principles depicted therein, are conceivable, insofar as they would typically occur to a person skilled in the art in the field of the invention.

[0023] It will be clear to those skilled in the art that the foregoing general description and the following detailed description are exemplary and explanatory of the invention and are not to be understood as a limitation thereof.

[0024] References to “an aspect”, “another aspect”, or similar phrases in this description mean that a particular feature, structure, or property described in connection with the embodiment is included in at least one embodiment of the present disclosure. Therefore, phrases such as “in one embodiment”, “in another embodiment”, and similar expressions in this description may, but do not necessarily, all refer to the same embodiment.

[0025] The terms "includes," "comprehensive," or similar expressions denote non-exclusive inclusion. Thus, a procedure or method containing a list of steps does not only include those steps but may also include further steps not explicitly listed or inherent in the procedure or method. Likewise, the statement "includes..." for one or more devices, subsystems, elements, structures, or components, without further limitations, does not preclude the existence of other devices, subsystems, elements, structures, or components.

[0026] Unless otherwise defined, all technical and scientific terms used herein have the same meanings generally known to those skilled in the art in the field to which this invention belongs. The systems, methods, and examples described herein serve only for illustration and are not to be understood as limiting.

[0027] Embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0028] The functional units described in this specification are referred to as devices. A device may be implemented in programmable hardware such as processors, digital signal processors, central processing units, FPGAs, PALs, PLDs, cloud processing systems, or similar. Devices may also be implemented in software for execution by various processor types. An identified device may contain executable code and, for example, comprise one or more physical or logical blocks of computer instructions, which may be organized as an object, procedure, function, or other construct. However, the executable files of an identified device need not be physically related; they may consist of different instructions stored in different locations that, when logically combined, constitute the device and fulfill its purpose.

[0029] The executable code of a device or module can consist of a single instruction or multiple instructions and can even extend across different code sections, applications, and storage media. Similarly, operational data within the device can be identified and represented, and can exist in any suitable form and be organized in any data structure. The operational data can be captured as a single data record or distributed across various storage media and may exist, at least partially, as electronic signals within a system or network.

[0030] References to “a selected embodiment”, “an embodiment”, or “an embodiment” in this description mean that a particular feature, structure, or property described in connection with the embodiment is included in at least one embodiment of the disclosed subject matter. Therefore, the phrases “a selected embodiment”, “in an embodiment”, or “in an embodiment” appearing at different points in this description do not necessarily refer to the same embodiment.

[0031] Furthermore, the described features, structures, or properties can be combined in one or more embodiments in any suitable manner. The following description contains numerous specific details to enable a comprehensive understanding of the embodiments of the disclosed subject matter. However, a person skilled in the art will recognize that the disclosed subject matter can also be realized without one or more of the specific details or with other methods, components, materials, etc. In other cases, known structures, materials, or processes are not presented or described in detail so as not to obscure aspects of the disclosed subject matter.

[0032] According to the exemplary embodiments, the disclosed computer programs or modules can be executed in a variety of ways, for example, as an application running in the memory of a device or as a hosted application running on a server and communicating with the device application or browser via various standard protocols such as TCP / IP, HTTP, XML, SOAP, REST, JSON, and other suitable protocols. The disclosed computer programs can be written in programming languages ​​that run either in the device's memory or on a hosted server, such as BASIC, COBOL, C, C++, Java, Pascal, or scripting languages ​​such as JavaScript, Python, Ruby, PHP, Perl, or other suitable programming languages.

[0033] Some of the described embodiments involve data transmission over a network, such as the transmission of various inputs or files. The network may include, for example, the internet, wide area networks (WANs), local area networks (LANs), analog or digital wired and wireless telephone networks (e.g., PSTN, ISDN, cellular networks, and xDSL), radio, television, cable, satellite, and / or other transmission or tunneling mechanisms for data. It may include multiple networks or subnetworks, each of which may, for example, have a wired or wireless data path. The network may include a circuit-switched voice network, a packet-switched data network, or another network for transmitting electronic data. For example, it may be based on the Internet Protocol (IP) or Asynchronous Transfer Mode (ATM) and support voice communication using VoIP, Voice over ATM, or similar protocols.In one embodiment, the network comprises a mobile network configured for the exchange of text or SMS messages.

[0034] Examples of networks include Personal Area Networks (PAN), Storage Area Networks (SAN), Home Area Networks (HAN), Campus Area Networks (CAN), Local Area Networks (LAN), Wide Area Networks (WAN), Metropolitan Area Networks (MAN), Virtual Private Networks (VPN), Enterprise Private Networks (EPN), the Internet, Global Area Networks (GAN), and so on.

[0035] Fig. Figure 1 shows a block diagram of a system (100) for estimating the effort for agile software development using multi-stage evolutionary stacking according to an embodiment of the present disclosure.

[0036] The system according to Fig. 1 comprises: a data acquisition module (102) that retrieves software effort datasets containing historical data on software development projects, including features and actual effort values, from one or more dataset repositories; a data preprocessing module (104) connected to the data acquisition module (102) that receives the software effort datasets from the data acquisition module, cleans the received data by removing inconsistencies with missing target values, and normalizes numerical input features to a common range; and a first-layer ensemble module (106) connected to the data preprocessing module (104) that comprises several heterogeneous base models, including a random forest model, a support vector regression model, and an extreme gradient boosting model. Each of these base models generates predictions based on the preprocessed datasets received from the data preprocessing module.System (100) further comprises a module (108) for optimizing genetic algorithms, which is connected to the ensemble module (106) of the first layer. This module is configured to encode weights as normalized real-valued vectors for each of the heterogeneous base learners, apply a fitness function to minimize the mean squared validation error and derive an optimal weight vector, assign optimized weights to the predictions of each of the heterogeneous base learners, and generate weighted predictions based on the optimized weights. System (100) further comprises a second-layer meta-learning module (110), which is connected to the genetic algorithm optimization module (108).This module is configured to receive the weighted predictions from module (108), process them using a deep multilayer perceptron network to learn complex patterns and nonlinear interactions, and generate a final software effort estimate. System (100) further includes an output processing module (112) connected to the second-level meta-learning module (110). This module receives the final software effort estimate and processes and visualizes the data to enhance user understanding. System (100) also includes a user interface (114) connected to the output processing module (112) to receive the processed final software effort estimate. The user interface (114) is configured to display the processed and visualized final software effort estimate.

[0037] In one embodiment, the data acquisition module (102) is further configured to retrieve the aforementioned data records from one or more data record repositories and to integrate data from multiple sources, including stored historical data records or real-time user input provided via the user interface. The data acquisition module is further configured to verify the completeness of the received software effort data records before transferring them to the data preprocessing module (104).

[0038] In one embodiment, the heterogeneous basic learners are selected to exhibit complementary bias-variance behaviors and are configured to handle nonlinear relationships in software effort datasets.

[0039] In one embodiment, the genetic algorithm optimization module (108) comprises: a chromosome representation mechanism that encodes the weights as a normalized real-valued vector, which includes a weight for the Random Forest model, a weight for the Support Vector Regression model, and a weight for the Extreme Gradient Boosting model.

[0040] In one embodiment, the genetic algorithm optimization module (108) is further configured to: iteratively develop weight combinations through selection, crossover, and mutation operations; evaluate the fitness of each weight combination based on the mean squared validation error; converge to the optimal weight vector that minimizes the prediction error; and dynamically adjust the weights based on data-driven evolutionary optimization.

[0041] In one embodiment, the genetic algorithm optimization module (108) is configured to function as an integral part of the decision-making process within a prediction pipeline, with the optimized weights serving as inputs for the second-level meta-learning module for hierarchical pattern learning.

[0042] In one embodiment, the output processing module (112) in conjunction with the user interface (114) is configured to provide a comparative analysis of the predicted versus the actual effort values, display visualizations of the predicted effort values, and generate a downloadable report with the final effort estimate for the software.

[0043] In one embodiment, the system (100) is configured to process a variety of heterogeneous datasets from agile software development and adaptively adjusts the contributions of the basic learners using the genetic algorithm optimization module to improve generalizability across different data conditions.

[0044] In one embodiment, the data acquisition module (102), the data preprocessing module (104), the first layer ensemble module (106), the genetic algorithm optimization module (108), the second layer meta-learning module (110), the output processing module (112) and the user interface (114) can be implemented in programmable hardware devices such as processors, digital signal processors, central processing units, field-programmable gate arrays, programmable array logic, programmable logic devices, cloud processing systems or the like.

[0045] The present invention relates to a system for estimating agile effort using machine learning with a genetic algorithm. The proposed MLES-based system optimizes the ensemble architecture itself, thus representing an improvement over previous models for estimating agile effort. The system uses a GA-based evolutionary optimization mechanism that dynamically adjusts the weights of the base learners within a multi-level architecture, thereby enabling adaptive, data-driven weight evolution.

[0046] The proposed Multi-Level Evolutionary Stacking (MLES) system integrates heterogeneous learning algorithms for agile effort estimation, with the MLES approach applying evolutionary optimization based on genetic algorithms (GA) at the ensemble composition level.

[0047] In this invention, a genetic algorithm (GA) dynamically develops optimal weight combinations of heterogeneous basic learning methods (RF, SVR, and XGBoost), while a deep MLP metal learner captures nonlinear and hierarchical dependencies between the GA-optimized predictions. The GA serves as the central, iterative learning component and not as an aid for feature or hyperparameter optimization. This evolutionary stacking strategy enables adaptive model integration across heterogeneous datasets, thus overcoming key limitations of static ensemble approaches.

[0048] Fig. Figure 2 shows a block diagram of the proposed effort estimation system using multi-stage evolutionary stacking according to an embodiment of the present disclosure.

[0049] In relation to Fig. 2. A genetic algorithm (GA) optimizes the stack layer architecture and model contributions within the system. The proposed two-layer learning approach delivers promising final results with accurate prediction of the effort required for agile software development.

[0050] As in Fig. As shown in Figure 2, historical data on software development projects, including features and actual effort, are obtained via software effort datasets. These datasets then undergo data preprocessing, which includes data cleaning, normalization, and splitting. Data cleaning removes inconsistencies, missing target values, or the effort required for cleaning. Additionally, datasets with missing or undefined input features are checked, and only complete datasets are retained for further processing. To avoid artificial bias, no statistical imputation is applied. This ensures valid supervised learning. Normalization scales all numerical input features to a common range. The data is split 80:20 between training and test data to evaluate model performance.Each dataset is independently split into 80 / 20 training and test data; no cross-dataset training or cross-validation is performed.

[0051] As in Fig.As shown in Figure 2, the system comprises a first layer of basic learners. To support ensemble techniques, the model uses multiple basic learners together with a genetic algorithm (GA) to optimize the results and assign weights. The first-layer approach is divided into two main phases. In the first phase, the basic model trains with SVR, RF, and XGBoost. In the second phase, regression is combined with genetic optimization. The system also includes evolutionary optimization using the genetic algorithm (GA), where the weights of the predicted machine learning (ML) algorithms are assigned by the GA. The predicted ML values ​​and the weights assigned by the GA are passed to the neural network in the second layer. The weights are encoded as normalized real-valued vectors using a chromosome representation mechanism.By using a fitness function, the validation MSE is minimized to determine the optimal weight vector. The system also includes a second layer of the meta-learner. This neural network, with multiple layers of interconnected nodes (neurons), learns complex patterns and makes predictions by recognizing complex relationships in the data. It uses the weighted predictions of the base models and learns to estimate the final software effort. The meta-learner architecture is specified as a deep MLP with an input layer of size 3, two hidden layers (64 and 32 neurons), a ReLU activation function, and an output layer with a linear activation function. Optimization was performed using the Adam optimizer (learning rate = 0.001). A deep MLP was chosen over linear regression as the meta-learner because it better captures nonlinear interactions between the GA-optimized outputs of the base models than linear models.Furthermore, Deep MLP offers an excellent ability to learn complex relationships between weighted predictions and contributes to improved performance stability. The test data were evaluated using the proposed MLES system and analyzed using various performance metrics. The performance of the proposed system is also assessed using MAE, MMRE, sMAPE, RMSE, and the R² score.

[0052] In one embodiment of the GA-based optimization, all parameters were selected through systematic experimental tuning. Various areas were tested to achieve a balance between exploration, convergence stability, and computational efficiency. The final values ​​were determined only after observing consistent performance, controlled diversity, and stable convergence over multiple test runs. This final configuration thus represents the best compromise between performance stability, computational effort, and robustness for 21 heterogeneous datasets used for effort estimation.

[0053] In one embodiment, the performance of the proposed system is evaluated using 21 datasets, which are used to assess effort estimation in software development using the system. The effectiveness of the effort estimation approach in software development, particularly in agile projects, is significantly influenced by dataset characteristics such as size, the number of descriptive attributes for the software project, the presence of missing values, and outliers.

[0054] Key performance indicators (KPIs) are used to verify overall performance and measure the results of proposed models compared to actual values. The KPIs MAE, MMRE, sMAPE, RMSE, and R are examples. 2Scores are used to evaluate the model's prediction accuracy. The mean absolute error (MAE) measures the average error magnitude across a set of predictions, without considering their direction. It is calculated as the average of the absolute differences between predicted and actual values. A lower MAE score indicates more accurate predictions. It provides a clear measure of the average prediction error and is helpful in interpreting data across different datasets with varying levels of effort. The mean magnitude-relevant error (MMRE)... evaluates the average relative error of all predictions, expressed as a percentage. It is calculated by taking the mean of the absolute percentage deviations between predicted and actual values.The mean relative error (MMRE) is frequently used in software project effort estimation to assess model performance; smaller values ​​reflect better performance. It expresses the error as a relative percentage, thus enabling a fair comparison between projects with different effort levels. Symmetric mean absolute percentage error (sMAPE) quantifies the average magnitude of a model's systematic error by determining the absolute deviation of each prediction and averaging these values. It accommodates a wide range of effort levels and avoids extreme percentage errors through symmetric normalization of the actual and predicted effort.Mean Squared Error (MSE) gives an impression of how far predictions deviate from actual values, but emphasizes larger errors more strongly by squaring the errors before averaging. It is well suited when large errors are particularly problematic, as the original effort units are preserved, thus making the order of magnitude of typical errors interpretable. The R. 2 The R-value is a statistical measure that indicates how well a model's predictions match actual values. It shows the proportion of the target variable's variability that can be explained by the model. The closer R-value is to the actual value, the better. 2 The closer it is to 1, the better the model explains the data. It measures how well a model explains the variance of the actual effort values ​​and thus indicates the overall predictive power of the estimation model.

[0055] Traditional machine learning techniques and individual basic learners for effort estimation are compared with the proposed MLES-based effort estimation system. The results show that the performance of the proposed system surpasses existing solutions. The proposed system is able to fine-tune weights by integrating generalization and emphasizing more accurate learners, which helps achieve this accuracy with generalization.

[0056] In one embodiment, a required public dataset is determined, validated based on the number of lines of code and actual effort, taking into account 21 datasets from established software engineering repositories. The MLES-based system uses this dataset as input and employs a multi-stage stacking strategy combined with evolutionary optimization to model the architecture and individual learning weights for improved predictive accuracy.

[0057] The proposed MLES system was compared to several standard models: LR, MLP, SVR, DT, RF, and XGBoost. The results were analyzed using common performance metrics such as MAE, MMRE, sMAPE, RMSE, and R. 2 analyzed. The results shown represent the average performance across all data sets.

[0058] The results show that the proposed model achieves lower MAE values ​​across almost all datasets. On complex datasets like Belady and Boehm, the system achieves MAE values ​​of 216.20 and 1092.34, respectively, significantly outperforming the next best models: SVR with 25077.40 on Belady and MLP with 57935.97 on Boehm. On the Kemerer dataset, the system achieves an MAE value of 94.37, which is significantly lower than all other approaches. A striking anomaly is observed in the result for the Desharnais dataset: here, it exhibits the greatest deviation and skewness, with an MAE value of 2983.01, which significantly exceeds the errors of SVR (60.53), RF (84.82), and XGBoost (72.15). The primary reason for this behavior lies in the characteristics of this dataset. As mentioned previously, the study has a small sample size and an extreme variance in the effort values, which leads to a distortion of the mean.The mean absolute error (MAE) across all datasets for the MLES system is 588.38. This significant reduction in absolute error underlines the performance and robustness of the model.

[0059] The proposed system was evaluated using various performance metrics. It achieves an average MMRE score of 1.12, which is lower than or comparable to current ensemble- and metaheuristic-based models. Many of these models exhibit MMRE scores between approximately 1.13 and 1.39 or are based on static weighting and grid-search optimization. The average MAE score of the proposed MLES system across all datasets is 588.38, which is lower than that of other models. This significant reduction demonstrates the robustness of the model. In direct comparison with literature values, the proposed MLES system offers a consistent reduction of 88% in MAE, 75% in sMAPE, and 93% in RMSE, and shows more stable RDS(on). 2-values ​​across different datasets. Unlike previous approaches, this model achieves improvement through GA-driven adaptive ensemble integration, rather than through manual or fixed model combinations.

[0060] The results showed an 88% reduction in MAE to an average value of 588.38. The MMRE dropped to 1.12, the lowest value of all comparison models. The RMSE reduction reached approximately 89.24%, with an average RMSE of 2584.69. The sMAPE reduction reached up to 75%, with an average sMAPE value of 0.4935, the lowest value of all models. The R 2Performance improved significantly, reaching a value close to zero, confirming its high variance resolution, especially with complex and noisy datasets. These results demonstrate that MLES significantly reduces the discrepancy between estimated and actual effort, effectively addressing a key challenge in agile project management. By improving the accuracy of effort estimation, the proposed MLES supports better resource allocation, sprint planning, effort distribution, and optimized risk management. Ultimately, this leads to higher project success rates and more predictable outcomes in agile project delivery.

[0061] The drawings and the preceding description illustrate embodiments. Those skilled in the art will recognize that one or more of the described elements can be combined to form a single functional element. Alternatively, certain elements can be divided into several functional elements. Elements of one embodiment can be added to another. For example, the process sequences described here can be modified and are not limited to the manner described herein. Furthermore, the actions of a flowchart need not be performed in the sequence shown; nor do all actions necessarily need to be carried out. Actions that do not depend on other actions can be performed in parallel with the other actions. The scope of protection of the embodiments is in no way limited by these specific examples. Numerous variations, whether explicitly stated in the description or not, such as...Differences in structure, dimensions, and materials are possible. The scope of protection of the embodiments is at least as comprehensive as described by the following claims.

[0062] The advantages, other benefits, and problem solutions have been described above with reference to specific embodiments. However, the advantages, benefits, problem solutions, and any components that can effect or enhance an advantage, benefit, or solution are not to be construed as critical, necessary, or essential features or components of the claims. REFERENCES 100 A system for estimating effort in agile software development using a multi-stage evolutionary stacking method. 102 Data Acquisition Module 104 Data preprocessing module 106 Ensemble module of the first layer 108 Optimization module for genetic algorithms 110 Meta-learning module of the second layer 112 Output processing module 114 User interface 202 data sets on software effort 204 Data preprocessing 206 training data points (80%) 208 Multi-stage evolutionary stacking 208a Basic learning methods (RF, SVR, XGBoost) 208b The predicted weights were assigned using a genetic algorithm (GA). 208c Deep MLP 208d Layer 1 208e Layer 2 210 test data (20%) 212 Model evaluated using test data 214 Performance Analysis

Claims

A system for effort estimation in agile software development using multi-stage evolutionary stacking, consisting of: a data acquisition module configured to retrieve software effort records from one or more dataset repositories containing historical data from software development projects with characteristics and actual effort values; a data preprocessing module operationally connected to the data acquisition module and configured to receive the aforementioned software effort records from the data acquisition module, clean the received data by removing inconsistencies with missing target values, and normalize numerical input characteristics to a common range;A first-layer ensemble module connected to the data preprocessing module, wherein the first-layer ensemble module comprises a variety of heterogeneous basic learners, including a random forest model, a support vector regression model, and an extreme gradient boosting model, which generate predictions from each of the heterogeneous basic learners using the preprocessed datasets received from the data preprocessing module; a genetic algorithm optimization module connected to the first-layer ensemble module and configured to: encode weights as a normalized real-valued vector for each of the heterogeneous basic learners; apply a fitness function to minimize the mean squared validation error and derive an optimal weight vector; assign optimized weights to the predictions of each of the heterogeneous basic learners; and generate weighted predictions based on the optimized weights;a second-level meta-learning module connected to the optimization module of the genetic algorithm, configured to receive the weighted predictions from the optimization module of the genetic algorithm, process the weighted predictions using a deep multilayer perceptron neural network to learn complex patterns and nonlinear interactions, and generate a final effort estimate for the software; an output processing module connected to the second-level meta-learning module, configured to receive the final effort estimate for the software, and process and visualize the data to enhance user understanding;and a user interface connected to the output processing module to receive the processed final effort estimate for the software, wherein the user interface is configured to display the processed and visualized final effort estimate for the software. System according to claim 1, wherein the data acquisition module is further configured to retrieve the data records from one or more data record stores and integrates data from multiple sources, including stored historical data records or real-time user input provided via the user interface, and wherein the data acquisition module is further configured to verify the completeness of the obtained software effort data records before transferring them to the data preprocessing module. System according to claim 1, wherein the plurality of heterogeneous basic learners is selected to provide complementary bias-variance behaviors and is configured to handle nonlinear relationships in software effort datasets. System according to claim 1, wherein the module for optimizing the genetic algorithm comprises: a chromosome representation mechanism encoding the weights as a normalized real-valued vector, comprising a weight for the Random Forest model, a weight for the Support Vector Regression model, and a weight for the Extreme Gradient Boosting model. System according to claim 1, wherein the module for optimizing the genetic algorithm is further configured to: iteratively develop weight combinations through selection, crossover, and mutation operations; evaluate the fitness of each weight combination based on the mean squared validation error; converge to the optimal weight vector that minimizes the prediction error; and dynamically adjust the weights based on data-driven evolutionary optimization. System according to claim 1, wherein the module for optimizing the genetic algorithm is configured as an integral part of a decision-making chain within a prediction pipeline and the optimized weights serve as inputs for the second-level meta-learning module for hierarchical pattern learning. System according to claim 1, wherein the output processing module in conjunction with the user interface is configured to provide a comparative analysis of the predicted effort versus the actual effort, displays visualizations of the predicted effort, and generates a downloadable report with the final effort estimate for the software. System according to claim 1, wherein the system is configured to process a variety of heterogeneous datasets from agile software development and adaptively adjusts the contributions of the basic learning models via the optimization module for genetic algorithms to improve generalizability under different data conditions.

Images