Inference in dynamic probabilistic models is a complex task involving expensive operations. In particular, for Hidden Markov Models, the whole state space has to be enumerated for advancing in time. Even states with negligible probabilities are considered, resulting in computational inefficiency and increased noise due to the propagation of unlikely probability mass. We propose to denoise the future and speed up inference by using only the top-p states, i.e., the most probable states with accumulated probability p. We show that the error introduced by using only the top-p states is bound by p and the so-called minimal mixing rate of the underlying model. Moreover, in our empirical evaluation, we show that we can expect speedups of at least an order of magnitude, while the error in terms of total variation distance is below 0.09.

In Dynamic Bayesian Networks, time is considered discrete: In medical applications, a time step can correspond to, for example, one day. Existing temporal inference algorithms process each time step sequentially, making long-term predictions computationally expensive. We present an exact, GPU-optimizable approach exploiting symmetries over time for prediction queries, which constructs a matrix for the underlying temporal process in a preprocessing step. Additionally, we construct a vector for each query capturing the probability distribution at the current time step. Then, we time-warp into the future by matrix exponentiation. In our empirical evaluation, we show an order of magnitude speedup over the interface algorithm. The work-heavy preprocessing step can be done offline, and the runtime of prediction queries is significantly reduced. Therefore, we can handle application problems that could not be handled efficiently before.

When allowing concurrent actions in Markov Decision Processes, whose state and action spaces grow exponentially in the number of objects, computing a policy becomes highly inefficient, as it requires enumerating the joint of the two spaces. For the case of indistinguishable objects, we present a first-order representation to tackle the exponential blow-up in the action and state spaces. We propose Foreplan, an efficient relational forward planner, which uses the first-order representation allowing to compute policies in space and time polynomially in the number of objects. Thus, Foreplan significantly increases the number of planning problems solvable in an exact manner in reasonable time, which we underscore with a theoretical analysis. To speed up computations even further, we also introduce an approximate version of Foreplan, including guarantees on the error. Further, we provide an empirical evaluation of both Foreplan versions, demonstrating a speedup of several orders of magnitude. For the approximate version of Foreplan, we also empirically show that the induced error is often negligible.

The growing practice of archiving research data in repositories reflects an upward trend. However, storing data in an RDR (Research Data Repository) does not guarantee that the archived data will always be readily reusable, even if this fulfils the FAIR (Findable, Accessible, Interoperable Reusable) principles. To ensure sustainable RDM (Research Data Management), archiving must consider the future potential for data reuse in a low-threshold fashion. In this article, we demonstrate the utilisation of straightforward methods to implement a so-called warm or hot archiving for research data within an RDR, as opposed to the conventional cold archiving approach. We explore the additional value of using research data in the humanities, emphasising the advantages of maintaining data accessibility and relevance over time. In the humanities, evaluating numerous data sets efficiently is crucial for current and future projects. Reviewing and evaluating relevance is important, particularly when dealing with a substantial number of data sets. Rapid evaluation facilitates profound decisions on the utility of the data for one’s ongoing or upcoming projects. For hot archiving, this means that in addition to the research data, the data should be available in a human-friendly way, i.e., a viewer application to visualise the data should be easily accessible. However, as rapid developments in the IT sector mean that after a few years, it cannot be guaranteed that these viewers or other tools will work, we also show how data can be viewed in a user-specific way via the RDR and how sustainable viewing can be integrated into the RDR. This article presents a generic approach to building sustainable viewers, which we call information systems, or transformer models on demand using data from pre-modern Arabic. In addition, we show that the easy-to-use chatbot ChatGPT can alternatively be context-specifically prepared to deliver more precise results and associated resources in the field of humanities. On the one hand, we have achieved a substantial reduction in the development time of an information system, from months to seconds, as well as the ability to fine-tune BERT (Bidirectional Encoder Representations from Transformers) models without specific knowledge in selecting models or tools. On the other hand, we have developed a chatbot that not only provides project-specific responses but also references the sources.

The runtime of graph algorithms such as depth-first search or Dijkstra's algorithm is dominated by the fact that all edges of the graph need to be processed at least once, leading to prohibitive runtimes for large, dense graphs. We introduce a simple data structure for storing graphs (and more general structures) in a compressed manner using directed acyclic graphs (dags). We then show that numerous standard graph problems can be solved in time linear in the size of the dag compression of a graph, rather than in the number of edges of the graph. Crucially, many dense graphs, including but not limited to graphs of bounded twinwidth, have a dag compression of size linear in the number of vertices rather than edges. This insight allows us to improve the previous best results for the runtime of standard algorithms from quasi-linear to linear for the large class of graphs of bounded twinwidth, which includes all cographs, graphs of bounded treewidth, or graphs of bounded cliquewidth.

Time in Bayesian Networks is concrete: In medical applications, a timestep can correspond to one second. To proceed in time, temporal inference algorithms answer conditional queries. But the interface algorithm simulates iteratively into the future making predictions costly and intractable for applications. We present an exact, GPU-optimizable approach exploiting symmetries over time during answering prediction queries by constructing a matrix for the underlying temporal process. Additionally, we construct a vector capturing the probability distribution at the current timestep. Then, we can time-warp into the future by matrix exponentiation. We show an order of magnitude speedup over the interface algorithm. The work-heavy preprocessing step can be done offline, and the runtime of prediction queries is significantly reduced. Now, we can handle application problems that could not be handled before.

As our contribution to the effort of developing methods to make robots more adaptive and robust to dynamic environments, we have proposed our method of ‘minimal surprise’ in a series of previous works. In a multi-robot setting, we use evolutionary computation to evolve pairs of artificial neural networks: an actor network to select motor speeds and a predictor network to predict future sensor input. By rewarding for prediction accuracy, we give robots an innate, task-independent motivation to behave in structured and thus, predictable ways. While we previously focused on feasibility studies using abstract simulations, we now present our first results using realistic robot simulations and first experiments with real robot hardware. In a centralized online and onboard evolution approach, we show that minimize surprise works effectively on Thymio II robots in an area cleaning scenario.

Real world scenarios can be captured with lifted probability distributions. However, distributions are usually encoded in a table or list, requiring an exponential number of values. Hence, we propose a method for extracting first-order formulas from probability distributions that require significantly less values by reducing the number of values in a distribution and then extracting, for each value, a logical formula to be further minimized. This reduction and minimization allows for increasing the sparsity in the encoding while also generalizing a given distribution. Our evaluation shows that sparsity can increase immensely by extracting a small set of short formulas while preserving core information.