In the past two decades, a range of model versioning workflows have been proposed. Standard workflows are based on three-way model merging, which allows reasoning on potentially conflicting changes in concurrently developed model versions. However, the considered conflicts that can be detected are mostly targeting the syntactic level of models, such as update/update or delete/usage conflicts. In contrast, unintended semantic inconsistencies often remain unnoticed as detection mechanisms lack the semantic awareness of the modeling language or modeled domain. The resolution of such conflicts remains a manual task. In this paper, we explore how Large Language Models (LLMs) can augment model versioning workflows by supporting conflict detection and resolution. In particular, we present an LLM-enhanced solution for detecting conflicts in the three-way model merging setting. Drawing on a collection of conflict types from prior literature, we demonstrate how an LLM assistant can 1) pinpoint conflicting changes and 2) provide resolution options with clear rationales and explanations of their implications. Our results indicate that the LLMs' access to a broad range of domains and modeling languages can help find and resolve complex versioning conflicts. Our implementation combines the industrial tool LemonTree for analyzing models and model changes, with a GPT-4o (LLM) assistant primed with relevant context to detect and resolve conflicts. We conclude by discussing directions for future research to improve model versioning workflows using LLMs.

Abstract Model-driven optimization allows to directly apply domain-specific modeling languages to define models which are subsequently optimized by applying a predefined set of model transformation rules. Objectives guide the optimization processes which can range from one single objective formulation resulting in one single solution to a set of objectives that necessitates the identification of a Pareto-optimal set of solutions. In recent years, a multitude of reinforcement learning approaches has been proposed that support both optimization cases and competitive results for various problem instances have been reported. However, their application to the field of model-driven optimization has not gained much attention yet, especially when compared to the extensive application of meta-heuristic search approaches such as genetic algorithms. Thus, there is a lack of knowledge about the applicability and performance of reinforcement learning for model-driven optimization. We therefore present in this paper a general framework for applying reinforcement learning to model-driven optimization problems. In particular, we show how a catalog of different reinforcement learning algorithms can be integrated with existing model-driven optimization approaches that use a transformation rule application encoding. We exemplify this integration by presenting a dedicated reinforcement learning extension for MOMoT. We build on this tool support and investigate several case studies for validating the applicability of reinforcement learning for model-driven optimization and compare the performance against a genetic algorithm. The results show clear advantages of using RL for single-objective problems, especially for cases where the transformation steps are highly dependent on each other. For multi-objective problems, the results are more diverse and case-specific, which further motivates the usage of model-driven optimization to utilize different approaches to find the best solutions.

For continuously checking and updating the virtual representation of a real system during operation, the continuous sensing and interpretation of raw sensor data is a must. The challenge is to bundle sensor value streams (e.g., from IoT networks) and aggregate them to a higher logical state level to enable process-oriented viewpoints and to handle uncertainties about sensor measurements and state realization precision. To address these uncertainties, so-called “tolerance ranges” must be defined in which logical states are detected during operation with acceptable deviations. Specifying such tolerance ranges manually is a time-consuming, error-prone task and often not feasible due to the huge associated value search space. To tackle this challenge, the problem is turned into an optimization problem in this paper. For this purpose, we present a framework based on meta-heuristic search that enables the automatic configuration of tolerance ranges based on available execution traces of multiple sensor value streams. An exploratory study evaluates the approach. For this purpose, we implemented a lab-sized demonstrator of a five-axis grip arm robot, which we continuously monitored during operation in a simulated environment. The evaluation shows the advantage of using meta-heuristic optimizers such as Harmony Search or Genetic Algorithm to identify stable tolerance ranges automatically for state detection at runtime.Note to Practitioners—Monitoring sensor values streams is nowadays a frequently employed technique in many automation domains. However, combining and mapping single value streams to higher-level state-based representations such as state machines or other design-time related models is a major challenge due to measurement and realization precision uncertainties. Thus, simply mapping monitored raw data to these design descriptions can lead to falsely identified or missed states. To improve this situation, we present an approach that provides a mechanism to continuously analyze data streams during operation by automatically finding appropriate tolerance ranges to detect realized system states. The approach uses a small set of annotated execution traces and meta-heuristic searchers to derive optimal tolerance ranges, which provide high correctness and completeness of the identified system states. This approach represents the basis for building a “vertical bridge” from the operation technology layer considering pure sensor data streams to the IT layer where state-based process views are provided to perform monitoring and analytics, e.g., by using process mining.

The increasing complexity of modern systems leads to an increasing amount of artifacts that are used along the model-based software and systems development lifecycle. This also includes model transformations, which serve for mapping models between representations, e.g., for verification and validation purposes. Model repositories manage this variety of artifacts and promote reusability, but should also enable the bundling of compatible artifacts. Therefore, model transformations should be reused and arranged into transformation chains to support more complex transformation scenarios. The resulting transformation should correspond to the user’s interest in terms of quality criteria such as model coverage, transformation coverage, and number of transformation steps, thus assembling such chains becomes a multi-objective problem. A novel multi-objective approach for exploring possible transformation chains residing in model repositories is presented. MOMoT, a model-driven optimization framework, is leveraged to explore the transformation space spanned by the repository. For demonstration, three differently populated repositories are considered. We have extended MOMoT with an exhaustive, multi-objective search that explores the entire model transformation space defined by graph transformation rules, allowing all possible transformation chains to be considered as solution. Accordingly, the optimal solutions were identified in the demonstration cases with negligible computation time. The approach assists modelers when there are multiple chains for transforming an input model to a specified output model to consider. Our evaluation shows that the approach elicits all legitimate transformation chains, thus enabling the modelers to consider trade-offs in view of multiple criteria selection.

Model-driven optimization has gained much interest in the last years which resulted in several dedicated extensions for in-place model transformation engines. The main idea is to exploit domain-specific languages to define models which are optimized by applying a set of model transformation rules. Objectives are guiding the optimization processes which are currently mostly realized by meta-heuristic searchers such as different kinds of Genetic Algorithms. However, meta-heuristic search approaches are currently challenged by reinforcement learning approaches for solving optimization problems. In this new ideas paper, we apply for the first time reinforcement learning for in-place model transformations. In particular, we extend an existing model-driven optimization approach with reinforcement learning techniques. We experiment with value-based and policy-based techniques. We investigate several case studies for validating the feasibility of using reinforcement learning for model-driven optimization and compare the performance against existing approaches. The initial evaluation shows promising results but also helped in identifying future research lines for the whole model transformation community.

An M/G/1 queue with server vacations and nongated time-limited service is analyzed. A functional equation which characterizes the system behavior is derived. The equation is solved by a numerical technique that approximates the unknown function by a weighted sum of Laguerre functions with unknown coefficients. The functional equation is transformed into a set of linear equations from which the coefficients can be computed. By the work-decomposition and Poisson arrivals see time averages (PASTA) properties, the average customer response time can be readily obtained for the case of exponential service time. Numerical examples are included to demonstrate the validity of the technique.>

An M/G/1 queue with server vacations and gated time-limited service is analyzed. At each visit, the server serves the queue up to a fixed amount of time. When the time expires or after all candidate customers have been served, whichever occurs first, the server takes a vacation. The service policy is gated, since only those customers present at the beginning of a server visit (poling instant) are candidates for service during that visit; subsequent arrivals are deferred until the next visit. A functional equation which characterizes the amount of work, U/sub p/, at a polling instant is derived. To solve the equation, a numerical technique is utilized in which the complementary cumulative function for U/sub p/ is closely approximated by a weighted sum of Laguerre functions with unknown coefficients. The equation is then transformed into a set of linear equations from which the coefficients can be computed. By the stochastic decomposition and Poisson-arrivals-see-time-averages properties, the average customer response time can be related to the average amount of work found by an arrival. Several numerical examples are included. The model studied is applicable to communication and computer systems where timers are used to allocate service to customers.>

A method for frame synchronization in a photonic network of time multiplexed space switches is presented. The method, which is based on a feedback scheme, is first introduced for frame synchronization in a single-switch configuration. The feedback scheme is then extended to both hierarchical (tree topology) and nonhierarchical (general topology) multiswitch configurations. Using this feedback scheme has the advantage of providing a mechanism for frame synchronization to be achieved initially, without the need for accurate transmission time measurements, and allowing synchronization to be maintained within very tight bounds. The methods presented require additional electronics at the switches, but electronics only for control purposes, not in the path of the information bits. However, optical delay elements are needed in the path of information bits on some links of the network. The proposed schemes also reduce the amount of synchronization equipment needed, compared to other methods.>

An analysis is conducted of an M/G/1 queue with server vacations and gated time-limited service. The authors derive a functional equation which characterizes the amount of work, U/sub p/, at the server's return from a vacation. To solve the equation, they use a numerical technique in which the complementary cumulative function for U/sub p/ is closely approximated by a weighted sum of Laguerre functions with unknown coefficients. The functional equation is transformed into a set of linear equations from which the coefficients can be computed. Using the work-decomposition and PASTA (Poisson Arrivals Sec Time Averages) properties, the average customer waiting time can be readily obtained. Several numerical examples are included to demonstrate the validity of the technique. The model studied is applicable to analyzing a specific recently proposed communication channel that alternately serves voice and data traffic, token-passing networks with token-holding timers, and other communication and computer systems where times are used to allocate service among multiple types of customers.>

The effects of unbalanced traffic on the performance of the multihop network are discussed. The basic question addressed is how performance might degrade as network size increases. Because unbalanced loads can occur in a multitude of ways, two separate approaches were taken to characterize the effects of imbalance. Present results show that the allowable throughput per user for realistic load patterns is reduced by a factor between about 0.3-0.5 from that predicted for uniform load. Contrary to original fears, this deloading factor is not very sensitive to network size. Although emphasis is on the performance of the multichannel multihop network, the methods presented are applicable to other communication networks which incorporate some type of hopping.>