Context: Artificial Intelligence (AI)-based software has gained increasing interest, especially in the last decade, due to advancements in underlying technologies and demands in varying business domains. With the proliferation to develop such software, there appears a need for developing methods and supporting tools/frameworks. Purpose: In this paper, we focus on tools/frameworks to automate AI-based software development process, from a holistic view. We review the scientific studies that were empirically validated and also evaluate their proposals with respect to basic characteristics including theme, research methods, types, domains, and a number of cases in empirical validations. Method: We elicit relevant studies (with the contribution type of tool or framework) from a larger set of primary studies identified by a systematic literature review on AI-based software development process. We select 14 primary studies in this context and analyze them with respect to the purposes of the proposals. Results: We review tools/frameworks that support AI-based software development process under four headings: software system development process, the development process of fair software, model development process, and model deployment and operation processes. We observe that domains of empirical validation are diverse while the number of empirical cases applied for validation is limited. Also, only half of the primary studies provide links to their proposals as open-source, which is very important for the repeatability of the empirical validations.
Almost every day, we hear about intelligent systems such as self-driving cars and unmanned aerial vehicles that come to life with artificial intelligence. In addition and in parallel to the era of digitalization, the demand for smart systems that analyze existing data and turn it into an advantage for institutions and customers in many business areas such as trade, health, finance, production, education, etc. is increasing. Accordingly, the growing scale of Internet-based systems and innovations in social platforms, autonomous systems and cloud computing, and IWSM-MENSURA '22: The Joint Conference of the 31th International Workshop on Software Measurement (IWSM) and the 16th International Conference on Software Process and Product Measurement (MENSURA) Sep 28-30, 2022, Izmir, Turkey * Corresponding author. tugba@cs.hacettepe.edu.tr (T. G. Erdogan); altunel@bilkent.edu.tr (H. Altunel); atarhan@cs.hacettepe.edu.tr (A. K. Tarhan) 0000-0003-1491-8739 (T. G. Erdogan); 0000-0003-1103-3644 (H. Altunel); 0000-0003-1466-9605 (A. K. Tarhan) the advancement of artificial intelligence (AI) and machine learning (ML) techniques in the last decade, have created a new specialty for the software industry: AI-based software system development.
Several studies have reported challenges about AI-based software development, more frequently for ML software [1,2,3] or deep learning (DL) software [4], and more specifically for development phases such as requirements [5] or testing [6]. In addition, large companies such as Google and Microsoft have shared their experiences of AI-based (or data-centric) software development in recent years. For example, Google reported that machine learning software used in everyday life will likely require ongoing maintenance costs [7]. Microsoft reported about the challenges AI teams face in managing data, component abstraction, and reusing software, noting that customizing and reusing AI-based software requires different methods and skills than those applied in traditional software development [8]. Consequently, development methods and supporting technologies must be adapted to meet the specific challenges of AI-based software engineering.
As a response to the need stated above, in this paper, we focus on and review in detail the studies in scientific (or white) literature, which propose the tools/frameworks to support the development process of AI-based software. We select 14 studies that have been empirically validated from a larger set of primary studies identified by a systematic literature review (SLR) on AI-based software development process. We review the studies by grouping them according to their purposes of proposals and evaluate their proposals with respect to basic characteristics including theme, research methods and types, domains and number of cases in empirical validations, contexts of empirical validations, and open links to the proposed tools/frameworks.
The rest of this paper is organized as follows. Section 2 provides a summary of the related studies. Section 3 overviews the study search and selection processes, and potential validity threats. Section 4 reviews the primary studies in detail, and Section 5 evaluates them with respect to basic characteristics of validation. Finally, Section 6 concludes the paper.
A list of related secondary studies, which include primary studies on tools/frameworks for AI-based software development, are provided in Table 1. As seen from the table, there is no study that solely investigates, reviews in detail, and evaluates the tools/frameworks which support the AI-based software development process. Therefore, ours is the first study that is intended for this specific purpose.
First, we have applied guidelines to conduct systematic literature reviews in software engineering [17], and have searched and identified 85 primary studies that propose development models or processes for AI-based software. In search and study selection, we have not focused on specific phases of software development such as requirements or design; rather, we took a holistic view for the software development process.
Related secondary studies Ref. Focus [9] This survey study presents a comprehensive overview of recent and cutting-edge AI software, including open-source frameworks and libraries, which support the implementation of ML and DL algorithms. However, it does not specifically focus on the development process of these algorithms. [10] This study mines both academic and grey literature, identifies 29 engineering best practices for ML applications, and conducts a survey among 300+ practitioners to determine the degree of adoption of these practices. Although it refers to a number of tools in both white and grey literature, its focus is not on tools/frameworks that support the development process, and it does not include a detailed review. [11] This survey describes major research efforts where ML systems have been deployed at the edge of computer networks, focusing on the operational aspects including compression techniques, tools, frameworks, and hardware used in successful applications of intelligent edge systems. Yet, it does not focus on tools/frameworks to support the development or deployment processes. [12] This study summarizes secondary data from non-grey and grey academic literature. It identifies, analyzes, and synthesizes the challenges of ML-enabled software development, and re-frames the traditional SE development process to engineer the ML software development process. However, it does not specifically review the primary studies that propose tools/frameworks to support the development process. [13] This study presents the results from a systematic mapping study on the life cycle of AI models. It yields 405 primary studies, mapped in 5 different research topics. It reports that only a minority of publications focus on data management and model production problems, and that more studies should address the AI life cycle from a holistic perspective. This study does not have a specific focus on tools/frameworks that support the life cycle of AI models. [3] This study identifies and synthesizes the current state of software engineering (SE) research for engineering ML systems by a systematic literature review of 141 primary studies. It highlights that SE aspects do not have a mature set of tools and techniques. It reviews in detail the primary studies under the knowledge areas of SWEBOK [14], including the ones that propose tools/frameworks in each area. However, only few of these primary studies support the development process of ML software. [15] The study surveys existing efforts for DevOps, and categorizes them according to the life cycle stages that they contribute. It outlines directions for future work in quality-aware DevOps, in particular at AI for DevOps and DevOps for AI software. In terms of tools/frameworks, it refers to only one primary study which proposes a software architecture solution that can ensure the continuous development of computer vision applications. [16] To synthesize knowledge on SE approaches for building, operating, and maintaining AI-based systems, this study presents results from a systematic mapping of 248 primary studies. It classifies their subjects to SWEBOK's knowledge areas, including the "Software Engineering Process" area and its supporting tools/frameworks. However, it only refers to these studies in the classification scheme and does not include a detailed review.
Research questions (RQs) of this review study are described below:
Search sentences that were run in digital libraries as of January 2022 are listed below:
1. ("artificial intelligence software" OR "AI software" OR "machine learning software" OR "ML software") AND ("development approach" OR "development process" OR "development model" OR "development life-cycle") 2. ("artificial intelligence applications" OR "AI applications" OR "machine learning applications" OR "ML applications") AND ("development approach" OR "development process" OR "development model" OR "development life-cycle") 3. ("artificial intelligence software" OR "AI software" OR "machine learning software" OR "ML software") AND ("software development" OR "application development") AND ("approach" OR "process" OR "model" OR "life-cycle") 4. ("artificial intelligence applications" OR "AI applications" OR "machine learning applications" OR "ML applications") AND ("software development" OR "application development") AND (approach OR process OR model OR life-cycle) 5. ("artificial intelligence software" OR "AI software" OR "machine learning software" OR "ML software") AND (maturity OR capability) AND ("software development" OR "software engineering") AND (approach OR process OR model OR life-cycle) 6. ("artificial intelligence applications" OR "AI applications" OR "machine learning applications" OR "ML applications") AND (maturity OR capability) AND ("software development" OR "software engineering") AND (approach OR process OR model OR life-cycle) 7. ("artificial intelligence software" OR "AI software" OR "machine learning software" OR "ML software") AND (challenges OR difficulties OR problems) AND ("software development" OR "software engineering") 8. ("artificial intelligence applications" OR "AI applications" OR "machine learning applications" OR "ML applications") AND (challenges OR difficulties OR problems) AND ("software development" OR "software engineering")
The number of studies retrieved from the digital libraries by these search sentences are given in Table 2. While identifying and selecting the primary studies, we have eliminated studies applying inclusion and exclusion criteria in the order of the digital libraries shown in the table. The inclusion criteria are: I1) papers focus on the holistic view of software development process rather than specific phases of software development; I2) papers are in the area of both software engineering and AI. The exclusion criteria are: E1) publication language being other than English; E2) full-text not being available; E3) books and theses; E4) duplicate results from different search methods, and E5) papers are explicitly short (e.g., less than six pages).
Until now, we have identified contributions made and research methods [18] employed in the primary studies, which are listed in Table 3. Full details can be found in the study's transparent online spreadsheet (https://tinyurl.com/ytnxhhuy). Here, we should note that we are still in the data extraction stage for the larger pool, and present the initial results from it on the basis of supporting tools/frameworks in this paper. (1) 5 (excluded in this paper),
In order to construct the study pool in line with the purpose of this paper, we have selected 14 primary studies that have a contribution type of tool or framework and a research method type of validation or evaluation. In validation research, techniques investigated are novel and have not yet been implemented in practice (used for example experiments, i.e. work done in the lab) while in evaluation research, techniques are implemented in practice (i.e. solution implementation) and an evaluation of the technique is conducted in terms of benefits and drawbacks (i.e. implementation evaluation) [19]. The reason for selecting a subset of primary studies with the stated characteristics is that, as we already mentioned in the introduction section, we would like to identify, review and evaluate the empirically validated tools/frameworks that support the development process of AI-based software.
A number of validity threats are of concern for this study as adopted from [20,21] in terms of construct, internal, external and conclusion validity. The guideline for conducting SLRs has been followed and a research protocol has been developed and reviewed to cope with construct validity. One threat of this type regarding the applied protocol can be excluding the studies for the specific phases of software development, such as requirements and design. However, we have done this purposefully, since the subject area is quite broad and we have considered the software development process as a whole. With this strategy applied, we have not only retrieved the studies that propose development models or processes as a primary contribution but also the studies that embed such contributions secondarily into their AI-based software development efforts. Another threat of construct validity is related to the research method (i.e. SLR) that we have employed in this study. Including the grey literature in addition to the white literature could have elicited further proposals of tools/frameworks for the AI-based software development process. We claim this threat so that it can be considered while reading our findings and also be addressed in future studies. In terms of internal validity, we have applied independent voting among authors for inclusion/exclusion of the studies and for quality assessment of initially included studies. Additionally, we have carried out pilot data extraction and have conducted regular meetings to ensure consistency and commonality during data extraction. Regarding external validity, we admit that the results we obtained are valid only for the primary study pool that we have created, and it cannot be generalized in other contexts. Finally, in terms of conclusion validity, we have made several meetings to analyze and synthesize the results of data extraction. Since we have used a standardized format (i.e. Google Sheet) in data analysis, we believe other researchers with the same study pool would come up with similar results.
In this section, we review the selected primary studies on tools/frameworks that support AI-based software development process by grouping the studies under four headings. We performed thematic synthesis [22] to group the studies as: software system development process, development process of fair software, model development process, and model deployment and operation processes. The studies falling under these headings are elaborated in the following subsections.
Full details with respect to the research questions raised can be found in the study's online spreadsheet (https://tinyurl.com/ytnxhhuy) in the following columns, respectively: RQ1: Type (AD); RQ2: Purpose (AO), theme (AR), domain (AV); RQ3: Link to download (AQ); RQ4: Validation type (Z, AA), research methods (AB-AK), context (AV).
The primary studies grouped under this heading are summarized in Table 4, each with its title, purpose and scope. As seen from the table, these three studies have different concerns, which indicates that there is a diversity of the studies in this field. The first one is related to health informatics, and we know there is a trend in applying AI techniques in this domain to reveal important insights about healthcare data and healthcare processes. The second one is about deep learning-based project development using the principles of MDE (Model Driven Engineering) to solve the challenges to build the deep learning-based project development, particularly for neural networks which have complicated architecture and dependencies. The third one is seeking solutions for the scalability and performance problems for the development of conversational AI software like mobile software, and proposes a framework inspiring Cell theory and BDI (Belief-Desire-Intention) software model. These proposals can be considered as example solutions to common problems in AI software development, and as the trending topics for the researchers.
Studies that support software system development process
Purpose and Scope P1 [23], 2020 -It investigates the interaction between software engineering and machine learning within the context of health systems.
-It proposes SEMLHI framework (the framework and methodology of software engineering for ML in health informatics) that includes four modules: software, machine learning, machine learning algorithms, and health informatics data.
-The SEMLHI methodology includes seven phases: designing, implementing, maintaining and defining workflows; structuring information; ensuring security and privacy; performance testing and evaluation; and releasing the software applications. -CellS framework has five important components: anima, delibera, cell, plan and ligand.
-It is used in a resources-constrained mobile system for conversational AI software like a software dialogue system.
The primary studies grouped under this heading are summarized in Table 5. It is apparent from the table that most of the studies focus on bias mitigation in ML software. Measurement of bias is handled in a maturity framework from the organizational level by applying ethical guidelines and standards, and fairness measurement is handled through a scalable measurement system for computing fairness metrics. It is obvious that ethical AI activities should be addressed with bias detection, bias/fair measurement, and bias mitigation methods starting from the initial stages of the AI development life cycle.
Studies that support development process of fair software
Purpose and Scope P3 [26], 2019 -It presents a maturity framework based on ethical principles and best practices to evaluate the organization capability in order to effectively govern AI bias.
-It provides a measurement of bias governance capability maturity of the organizations by separating the scoring in two sections as "consideration" and "action". P8 [27], 2021 -It introduces Fairea, a model behaviour mutation approach for benchmarking and quantitatively evaluating bias mitigation methods for ML software.
-It includes three bias mitigation approaches: pre-processing, in-processing and postprocessing.
-There are three primary steps: baseline creation with model behaviour mutation, bias mitigation effectiveness region division, and quantitative evaluation of trade-off effectiveness. P9 [28], 2020 -It proposes a method named Fairway for bias detection and bias mitigation in binary classification models to remove ethical bias from training data and trained models.
-This handy tool combines two bias mitigation approaches: pre-processing (before model training) and in-processing (while model training). P11 [29], 2020 -It presents LinkedIn Fairness Toolkit (LiFT), a framework for scalable computation of fairness metrics as part of ML systems.
-It highlights the key requirements in deployed settings, and presents the design of a fairness measurement system. -LiFT comprises bias measurement and mitigation components that can be integrated into different stages of an ML training and serving system.
The primary studies grouped under this heading are summarized in Table 6. As seen from the table, even though introduced models have different specific targets, they all have the common focus on data contribution in model development. The first model focuses on the data scientists' cognitive workflows that include data collection. The second model introduces a validation framework for data. The last model provides an open source library with different data modalities.
The primary studies grouped under this heading are summarized Table 7. It is apparent from the table that all studies for model deployment and operation take software, data and its machine learning models into account together. That means pipeline and deployment of software are not independent of data and machine learning models. In addition to that, the use cases are limited and only selected models are used. They need to be applied in different domains with real use cases to get wider insight for further usage areas.
After grouping and reviewing the primary studies in the previous section, we further evaluate their proposals with respect to basic characteristics of empirical validations, which include research method and type, domain and number of empirical cases, context of empirical cases, and link to proposed tool/framework. Table 8 presents these characteristics for all the included primary studies in columns.
As seen from the Table 8, the evaluation type of research ( 8) is slightly more than validation type (6) in the primary studies. Case study research (in eight studies) and lab experiments (in six studies) are the most frequently used research methods. While domains of empirical validation are diverse, the number of empirical cases applied for validation is limited. Contexts of empirical validation are various in alignment with the diversity of the domains. Finally, we see that only half (7) of the primary studies provide links to their proposals as open source, which is very important for the repeatability of the empirical validations as well as for further validation studies by other researchers or practitioners. -The infrastructure is CHASE-CI for managing fast GPU appliances for machine learning and storage managed through Kubernetes on the high-speed.
-It presents a software containerization approach and libraries for turning into a distributed computer for big data analysis.
-It explains a use case in earth science phenomena with object segmentation workflow.
In terms of the quantity of the cases included in empirical validation, the following studies come to the fore, only the first one employing validation type research: P2 [33] (for model deployment and operation processes) with seven cases from software intensive embedded system domain, P9 [28] (for development process of fair software) with five cases regarding five datasets used from UC Irvine ML Repository, and P13 [32] (for model development process) with five cases from mixed domains including telemetry, healthcare, energy, water supply, and data science.
In this study, we reviewed tools/frameworks within white literature in a systematic way for tools/frameworks that support the AI-based software development process from a holistic perspective. We selected and analyzed 14 primary studies and we grouped them under four headings: software system development process, the development process of fair software, model development process, and model deployment and operation processes. Within these four groups, the scopes of the studies are seen as diversified, and data and model dependency of the tools or frameworks is observed such as P1 proposing a framework for health informatics whereas P5 is focusing on mobile systems. Additionally, data and machine learning models have effects on the software development process, and only half of the studies include the links to their proposals. On the other hand, more than half of the studies are presented with evaluations while the rest of them use validation type research.
These findings underline the fact that the AI-based software development process is an emerging field of research and needs further investigation. Furthermore, some of the proposed tools or frameworks need to be sharpened and to be validated in different domains in various industrial contexts. For researchers and practitioners, the next step can be the comparison of new tools/frameworks with existing models in industry in terms of performance and quality metrics. Additionally, a comparison of the tools/frameworks that support traditional software development processes may provide valuable findings in the maturation of the overall software development in the AI/ML area. As the last point, practitioners in industry can test the seven of the tools/frameworks with their access links given in Table 8, evaluate their real-life performances, and share their results publicly as the feedback to the community as well as the researchers who introduced them.