Genetic programming (GP) produces solutions that, while often less performant than neural networks, are uniquely amenable to analysis. In domains like game-playing, this explainability could reveal why certain policies succeed or fail—but such insights are rarely leveraged to improve the algorithms themselves. Here, we analyze a GP algorithm for 2048, testing whether structural, behavioral, or semantic metrics of evolved policies correlate with performance. While structural metrics showed no predictive power and behavioral features yielded ambiguous results, we identified a semantic feature that correlated with policy quality. Using this insight, we designed a mutation operator that improved performance. Though modest, this improvement suggests that explainability metrics can guide operator design, even when broader explanatory goals remain unmet. More broadly, our work highlights a potential advantage of GP: its solutions may be analyzable in ways that opaque methods like neural networks are not.
Optimisation research requires possibilities to test and evaluate new algorithms and algorithm adaptions. Two interesting problem classes—relevant to society—that covers shortest path problems and resource allocation in a very dynamic environment, with the possibility of running endlessly, are ride-hailing and ride-sharing. This paper presents an extensible simulation framework for ride-hailing and ride-sharing problems that allows researchers to create solutions in any program language, provides a graphical user-interface, and automatic evaluation.
Modern manufacturing needs to be flexible and resilient, thus requiring dynamic production scheduling that considers different capabilities and costs of available workstations and robots, and optimizes multiple-objectives under varying operation conditions and changing goals. We present FactorySim, a simulation testbed for a modern dynamic factory that supports a) testing different scheduling and resilience strategies to cope with varying requirements of orders, availability of work stations and changing importance of objectives, and b) manual inspection through an interactive GUI. FactorySim allows researchers to deploy their own self-adaptive controllers at the task scheduling level and at the goal management level; the simulation setting can be customized easily, and the modularity of the architecture enables future extensions. We provide two adaptive schedulers and an example scenario in the exemplar.
GECCO’26 Competition poster of the Explainable Ride-Sharing Optimisation for Sustainable Traffic Organisation competition
The Ride-Hailing Problem is an online optimization challenge that coordinates a fleet of vehicles to serve requests revealed over time. The problem is NP-hard, and many solution approaches use stochastic or machine-learning based algorithms, which perform well but are difficult to interpret. We study the trade-offs between performance and explainability. Our results show that greedy assignment consistently outperforms more complex methods, indicating that explainability need not come at the cost of efficiency. We also find that the effectiveness of repositioning strategies strongly depends on demand patterns, highlighting the value of context-aware policies.
Hybrid systems are used to model cyber-physical systems. They combine continuous dynamics with discrete switching behavior. The manual identification of hybrid system models is time-consuming and error-prone, motivating data-driven approaches to hybrid system identification. Existing approaches focus primarily on learning the continuous dynamics and discrete modes of a hybrid automaton. For the guard conditions between modes, they apply probabilistic methods, which lack interpretable and verifiable representations. We propose the usage of signal temporal logic (STL) for learning guards. The advantages of our approach are twofold: it introduces human-readable and verifiable descriptions of transition logic and allows integrating prior knowledge by specifying templates for expected guard structures. Empirical results show that the accuracy of our approach is comparable to existing decision tree-based methods with respect to reconstructing system behavior. The learned guards are more expressive and interpretable. By this, we advance hybrid system identification towards trustworthy and explainable data-driven modeling.
With Virtual Prototypes (VPs), it is possible to significantly improve the debugging process during system design as the interaction of components can be analyzed and internal variable values can be accessed. But in case of detecting a fault, it remains challenging to determine its specific cause, as the load of available information can be overwhelming. To address this issue, we propose to monitor a VP at runtime to extract an under-approximating formal model of the system behavior, and derive causal explanations by model checking. For this, we determine the sufficient condition of a failure based on the order and context of events. This is demonstrated using an abstracted controller for a wind turbine, implemented using a RISC-V VP.
Modern interacting digital systems are becoming increasingly complex, making it difficult to ensure their actual behavior aligns with design-time expectations, particularly in uncertain or dynamic environments, even when specifications are correct. This misalignment affects system scalability, reliability, and increases maintenance costs. We introduce a conceptual framework for identifying and self-explaining mismatches between expected and observed system behavior, together with an algorithm that generates explanations and case studies that apply the conceptual framework for explanation generation in an interacting digital systems setting.
Abstractions of signals help to avoid cognitive and storage overload, especially in systems with many signals such as cyber-physical systems (CPS). For assessing trends and reconstructing behavior, it often suffices to have a rough understanding how signals evolve, e.g., whether their behavior is monotonic or periodic. This work provides a configurable abstraction of signal behavior, where signals are described by a sequence of oscillation and linear patterns. We formalize templates of both behaviors in parameterized signal temporal logic (PSTL) and provide an algorithm that abstracts signals in terms of those patterns. The templates are configurable such that they allow for choosing the level of abstraction, e.g., by limiting the approximation error or the minimal oscillation frequency. For segmentation of the signal, we solve an optimization problem using a modified version of TeLEx. The evaluation demonstrates that configuration of the patterns is suitable to define the level of abstraction. Further, on control output from the ARCH wind turbine benchmark and flow data from medical ventilation, it demonstrates that the abstraction method can be applied to real-world signals.
Neural networks (NNs) have great potential to improve individualization of medicine, e.g., through analysis of signals. However, they are generally not interpretable. Understanding NN decisions is crucial, especially in safety-critical domains such as medicine. This work presents a new method to provide local explanations for classifications of signals made by NNs. Our method extends the Sig-LIME explanation method from one-dimensional signals to multidimensional signals by introducing new perturbation techniques. We evaluate the proposed method on an NN that classifies the positive end-expiratory pressure (PEEP) applied by a ventilator. The evaluation shows that the generated explanations are plausible, stable and concise.
Goal management has been recognized as a core challenge in the development of true autonomous self-adaptive systems. Unlike behavioral adaptations, which can typically be justified by showing how a new behavior better aligns with a given goal, goal adaptations are much more difficult to design and even more so to justify. In this paper we investigate the goal management problem and the developments necessary towards achieving true self-autonomy at this level, and, we advocate for the use of decision theory tools in tackling this problem. More specifically we (i) describe a decision theory compatible model for the goal management problem, and we (ii) discuss six key challenges that we argue must be tackled in order to achieve true autonomous goal management. Our decision theory compatible model yields a new utility characterization theorem at an appropriate resolution to allow for algorithmic analysis, and facilitates further mathematical formulation of the challenges identified.
@inproceedings{AMF:2025,
language = {USenglish},
author = {Alkhiyami, Mohammad and Martino, Gianluca and Fey, Goerschwin},
title = {Explaining Mismatches in Expected versus Perceived Behavior for Interacting Digital Systems},
booktitle = {Explanations with Constraints and Satisfiability (ExCoS)},
year = {2025}
}