Robert Feldt's research while affiliated with Chalmers University of Technology and other places

... Enoiu, et al. [41] present a software testing cycle that is centered around their observation that test design and execution can be viewed as a classical problem solving process like the Plan Do Check Act cycle. Their study recognizes the need to explore how testers create test cases in future works. ...

... Many of the safety problems studied today are caused by algorithms exploiting flaws and unintended loopholes in the design of an objective function. Examples include agents finding and exploiting ways to cheat in games [3], agents exploiting bugs in simulators [14], and supervised learning agents that can be fooled as they have learned other patterns in their data than we expected [24]. As AI algorithms become smarter they might even find other exploiting methods that are not reliant on weaknesses in the objective function design [7]. ...

... In earlier work (Lee et al., 2019) we presented the basic technique and empirically evaluated its use for producing backward slices. This paper extends our previous work in three main ways. ...

... Detecting DNN bugs is not easier than detecting bugs in traditional programs, i.e., programs without any data-driven component in them, as DNNs depend on the properties of the training data and numerous hyperparameters [8]. Mitigating DNN bugs has been the subject of fervent research in recent years, and various techniques have been proposed for testing [9], [10], fault localization [11], [12], and repair [13], [14] of DNN models. ...

... While Type-I and Type-III projects are developed, respectively, by a small number of large open-source projects and many individual software organizations, a substantial proportion of Type-II projects, i.e., ML pipelines, are produced by researchers as a by-product of a published paper [5]. This is because most of the major research venues, both in the AI (e.g., AAAI [6], NeurIPS [7], ICML [8]) and software engineering domains (e.g., ICSE [9], EMSE [10], TSE [11]), have strict submission requirements to facilitate open science and reproducibility of research results, which requires submission of datasets and code along with their papers. Typically, researchers would upload a preprint of their work on a repository like ArXiv, including a GitHub repository with the pipeline code and (potentially) a labeled data set to train the pipeline on. ...

... Upcycling materials are project materials used as materials for upcycling. As Figure 3 shows, a wide variety of project materials can be used, including the entire project, source code files, functions, application programming interfaces (APIs), test cases, and project design documents (e.g., Unified Modeling Language (UML) diagrams and data model designs) [42][43][44]. Limiting the types of upcyclable materials that can be accumulated in the knowledge base risks narrowing the possibilities for upcycling. Hence, the system aims to accumulate all kinds of project materials regardless of type or format. ...

... We carried out an evaluation with students first because authors such as [53], [58] highlight the idea that the level of internal validity in experiments with students may be similar to that with practitioners (or even higher if there is prior training in the skills being evaluated). All the participants initially had similar knowledge of GSD and DSD, something about which they were asked at the beginning of the Pre-Test, and only a few students had previous experience in these types of contexts. ...

... Moreover, our work demonstrates that external stakeholders can get involved in many ways. In the future, it would be important to explore this in more detail, and evaluate the consequences as demonstrated by Steghöfer et al. (2018). ...

... For example, future works could use the well-formedness rules proposed in Section 4.6 to search for optimal refinements of security policies automatically. Similar to a proposal by Scandariato et al.[SHF18], such an approach could be used to optimize the security arrangements made by software architects, enabling design space exploration on the basis of search-based MDE [BSA17]. For example, optimizing refinements could ease the integration of third-party or legacy components with fixed or even unknown security policies, minimize the verification effort by reducing the number of policies that must be actively verified, or enforce the principle of least privilege by preventing policies from being more permissive than necessary. ...

... If the agent's reward is based on how much dirt it observes, it might find ways of reducing its vision, and thus not observe any dirt at all. Reward gaming has frequently been reported among researches [1,15], and is usually fixed with a more robust objective function. ...