While in traditional computer game playing agents were designed solely for the purpose of playing one single game, General Game Playing is concerned with agents capable of playing classes of games. Given the game's rules and a few minutes time, the agent is supposed to play any game of the class and eventually win it. Since the game is unknown beforehand, previously optimized data structures or human-provided features are not applicable. Instead, the agent must derive a strategy on its own. One approach to obtain such a strategy is to analyze the game rules and create a state evaluation function that can be subsequently used to direct the agent to promising states in the match. In this thesis we will discuss existing methods and present a general approach on how to construct such an evaluation function. Each topic is discussed in a modular fashion and evaluated along the lines of quality and efficiency, resulting in a strong agent.

Der Artikel beschreibt ein Vorgehen für transdisziplinäre Forschungsteams, die auf Basis der integrativen Rahmenkonzeption „Mensch-Cyber-Technik-System“ (Human-Cyber-Physical System) gemeinsame Forschungsfragestellungen bearbeiten möchten. Es wird eine Systematik skizziert, die am Beispiel versteh- und erklärbarer Künstlicher Intelligenz für die ausgewählte Applikation der „Modellierung von menschlichem Expertenwissen über die Elastizität von Materialien“ schrittweise erläutert wird.

This thesis focuses on improving the SAT solving technology. The improvements focus on two major subjects: sequential SAT solving and parallel SAT solving. To better understand sequential SAT algorithms, the abstract reduction system Generic CDCL is introduced. With Generic CDCL, the soundness of solving techniques can be modeled. Next, the conflict driven clause learning algorithm is extended with the three techniques local look-ahead, local probing and all UIP learning that allow more global reasoning during search. These techniques improve the performance of the sequential SAT solver Riss. Then, the formula simplification techniques bounded variable addition, covered literal elimination and an advanced cardinality constraint extraction are introduced. By using these techniques, the reasoning of the overall SAT solving tool chain becomes stronger than plain resolution. When using these three techniques in the formula simplification tool Coprocessor before using Riss to solve a formula, the performance can be improved further. Due to the increasing number of cores in CPUs, the scalable parallel SAT solving approach iterative partitioning has been implemented in Pcasso for the multi-core architecture. Related work on parallel SAT solving has been studied to extract main ideas that can improve Pcasso. Besides parallel formula simplification with bounded variable elimination, the major extension is the extended clause sharing level based clause tagging, which builds the basis for conflict driven node killing. The latter allows to better identify unsatisfiable search space partitions. Another improvement is to combine scattering and look-ahead as a superior search space partitioning function. In combination with Coprocessor, the introduced extensions increase the performance of the parallel solver Pcasso. The implemented system turns out to be scalable for the multi-core architecture. Hence iterative partitioning is interesting for future parallel SAT solvers. The implemented solvers participated in international SAT competitions. In 2013 and 2014 Pcasso showed a good performance. Riss in combination with Copro- cessor won several first, second and third prices, including two Kurt-Gödel-Medals. Hence, the introduced algorithms improved modern SAT solving technology.

While in traditional computer game playing agents were designed solely for the purpose of playing one single game, General Game Playing is concerned with agents capable of playing classes of games. Given the game's rules and a few minutes time, the agent is supposed to play any game of the class and eventually win it. Since the game is unknown beforehand, previously optimized data structures or human-provided features are not applicable. Instead, the agent must derive a strategy on its own. One approach to obtain such a strategy is to analyze the game rules and create a state evaluation function that can be subsequently used to direct the agent to promising states in the match. In this thesis we will discuss existing methods and present a general approach on how to construct such an evaluation function. Each topic is discussed in a modular fashion and evaluated along the lines of quality and efficiency, resulting in a strong agent.:Introduction Game Playing Evaluation Functions I - Aggregation Evaluation Functions II - Features General Evaluation Related Work Discussion

Von Mittwoch, den 21.04.2021, bis Freitag, den 23.04.2021, richtete das Archiv für deutsche Polarforschung (AdP) des Alfred-Wegener-Instituts (AWI), Helmholtz-Zentrum für Polar- und Meeresforschung, und der Verband deutscher Archivarinnen und Archivare (VdA) die Frühjahrstagung der Fachgruppe 8 des VdA aus. Diese Tagung war bereits im Frühjahr 2020 geplant und musste dann abgesagt werden. Im Jahr 2021 wurde die Tagung von vornherein als digitale Tagung geplant. Neben dem Tagungsprogramm wurde ein digitaler Pausenraum unter Wonder.me eingerichtet, in dem sich die Tagungsteilnehmer zwanglos zum kollegialen Gespräch treffen konnten. Die Tagung widmete sich dem Thema 'Digitalisierung' in seiner ganzen Breite - natürlich mit Bezug zu den Archiven in unserer Fachgruppe. Unter dieser Prämisse behandeln wir in Vorträgen, Workshops, Diskussionen und Führungen verschiedene Aspekte des Themas. Dazu gehören: Erhebung, Verarbeitung, Speicherung und Aufbewahrung von Forschungsdaten, Schriftgutverwaltung im Rahmen von Dokumentenmanagementsystemen, die Einrichtung von digitalen Archiven, Nutzung von Archiven im Rahmen von Portalen, rechtliche Aspekte des Digitalisierungsprozesses im Archiv und zukünftige Einsatzmöglichkeiten von Künstlicher Intelligenz in den digitalen Workflows eines Archivs. / From Wednesday, 21 April 2021, to Friday, 23 April 2021, the Archive for German Polar Research (AdP) of the Alfred Wegener Institute (AWI), Helmholtz Centre for Polar and Marine Research, and the Association of German Archivists (VdA) hosted the spring meeting of Division 8 of the VdA. This meeting was already planned for spring 2020 and then had to be cancelled. In 2021, the conference was planned from the outset as a digital conference. In addition to the conference programme, a digital break room was set up under Wonder.me, where conference participants could meet informally for collegial discussions. The conference was dedicated to the topic of 'digitisation' in all its breadth - naturally with reference to the archives in our section. Under this premise, we dealt with various aspects of the topic in lectures, workshops, discussions and guided tours. These include: collection, processing, storage and preservation of research data, records management in the context of document management systems, the establishment of digital archives, use of archives in the context of portals, legal aspects of the digitisation process in the archive and future possibilities of using artificial intelligence (AI) in the digital workflows of an archive.

The task of 6D object pose estimation, i.e. of estimating an object position (three degrees of freedom) and orientation (three degrees of freedom) from images is an essential building block of many modern applications, such as robotic grasping, autonomous driving, or augmented reality. Automatic pose estimation systems have to overcome a variety of visual ambiguities, including texture-less objects, clutter, and occlusion. Since many applications demand real time performance the efficient use of computational resources is an additional challenge. In this thesis, we will take a probabilistic stance on trying to overcome said issues. We build on a highly successful automatic pose estimation framework based on predicting pixel-wise correspondences between the camera coordinate system and the local coordinate system of the object. These dense correspondences are used to generate a pool of hypotheses, which in turn serve as a starting point in a final search procedure. We will present three systems that each use probabilistic modeling and sampling to improve upon different aspects of the framework. The goal of the first system, System I, is to enable pose tracking, i.e. estimating the pose of an object in a sequence of frames instead of a single image. By including information from previous frames tracking systems can resolve many visual ambiguities and reduce computation time. System I is a particle filter (PF) approach. The PF represents its belief about the pose in each frame by propagating a set of samples through time. Our system uses the process of hypothesis generation from the original framework as part of a proposal distribution that efficiently concentrates samples in the appropriate areas. In System II, we focus on the problem of evaluating the quality of pose hypotheses. This task plays an essential role in the final search procedure of the original framework. We use a convolutional neural network (CNN) to assess the quality of an hypothesis by comparing rendered and observed images. To train the CNN we view it as part of an energy-based probability distribution in pose space. This probabilistic perspective allows us to train the system under the maximum likelihood paradigm. We use a sampling approach to approximate the required gradients. The resulting system for pose estimation yields superior results in particular for highly occluded objects. In System III, we take the idea of machine learning a step further. Instead of learning to predict an hypothesis quality measure, to be used in a search procedure, we present a way of learning the search procedure itself. We train a reinforcement learning (RL) agent, termed PoseAgent, to steer the search process and make optimal use of a given computational budget. PoseAgent dynamically decides which hypothesis should be refined next, and which one should ultimately be output as final estimate. Since the search procedure includes discrete non-differentiable choices, training of the system via gradient descent is not easily possible. To solve the problem, we model behavior of PoseAgent as non-deterministic stochastic policy, which is ultimately governed by a CNN. This allows us to use a sampling-based stochastic policy gradient training procedure. We believe that some of the ideas developed in this thesis, such as the sampling-driven probabilistically motivated training of a CNN for the comparison of images or the search procedure implemented by PoseAgent have the potential to be applied in fields beyond pose estimation as well.

This work shows the application of artificial intelligence (AI) on invasion game tracking data to realize a fast (sub-second) and adaptable search engine for sports scenes, scene ratings based on machine learning (ML) and computer-generated solutions using reinforcement learning (RL). We provide research results for all three areas. Benefits are expected for accelerated video analysis at professional sports clubs. / Diese Arbeit zeigt die Anwendung von künstlicher Intelligenz (KI) auf Invasionsspielverfolgungsdaten, um eine schnelle (unter einer Sekunde) und anpassungsfähige Suchmaschine für Sportszenen zu realisieren, Szenenbewertungen auf der Grundlage von maschinellem Lernen (ML) und computergenerierte Lösungen unter Verwendung von Verstärkungslernen (RL). Wir stellen Forschungsergebnisse für alle drei Bereiche vor. Es werden Vorteile für eine beschleunigte Videoanalyse in Profisportvereinen erwartet.

Für mobile autonome Roboter ist ein solides Modell der Umwelt eine wichtige Voraussetzung um die richtigen Entscheidungen zu treffen. Die gängigen existierenden Verfahren zur Weltmodellierung basieren auf dem Bayes-Filter und verarbeiten Informationen mit Hidden Markov Modellen. Dabei wird der geschätzte Zustand der Welt (Belief) iterativ aktualisiert, indem abwechselnd Sensordaten und das Wissen über die ausgeführten Aktionen des Roboters integriert werden; alle Informationen aus der Vergangenheit sind im Belief integriert. Wenn Sensordaten nur einen geringen Informationsgehalt haben, wie zum Beispiel Peilungsmessungen, kommen sowohl parametrische Filter (z.B. Kalman-Filter) als auch nicht-parametrische Filter (z.B. Partikel-Filter) schnell an ihre Grenzen. Das Problem ist dabei die Repräsentation des Beliefs. Es kann zum Beispiel sein, dass die gaußschen Modelle beim Kalman-Filter nicht ausreichen oder Partikel-Filter so viele Partikel benötigen, dass die Rechendauer zu groß wird. In dieser Dissertation stelle ich ein neues Verfahren zur Weltmodellierung vor, das Informationen nicht sofort integriert, sondern erst bei Bedarf kombiniert. Das Verfahren wird exemplarisch auf verschiedene Anwendungsfälle aus dem RoboCup (autonome Roboter spielen Fußball) angewendet. Es wird gezeigt, wie vierbeinige und humanoide Roboter ihre Position und Ausrichtung auf einem Spielfeld sehr präzise bestimmen können. Grundlage für die Lokalisierung sind bildbasierte Peilungsmessungen zu Objekten. Für die Roboter-Ausrichtung sind dabei Feldlinien eine wichtige Informationsquelle. In dieser Dissertation wird ein Verfahren zur Erkennung von Feldlinien in Kamerabildern vorgestellt, das ohne Kalibrierung auskommt und sehr gute Resultate liefert, auch wenn es starke Schatten und Verdeckungen im Bild gibt. / For autonomous mobile robots, a solid world model is an important prerequisite for decision making. Current state estimation techniques are based on Hidden Markov Models and Bayesian filtering. These methods estimate the state of the world (belief) in an iterative manner. Data obtained from perceptions and actions is accumulated in the belief which can be represented parametrically (like in Kalman filters) or non-parametrically (like in particle filters). When the sensor''s information gain is low, as in the case of bearing-only measurements, the representation of the belief can be challenging. For instance, a Kalman filter''s Gaussian models might not be sufficient or a particle filter might need an unreasonable number of particles. In this thesis, I introduce a new state estimation method which doesn''t accumulate information in a belief. Instead, perceptions and actions are stored in a memory. Based on this, the state is calculated when needed. The system has a particular advantage when processing sparse information. This thesis presents how the memory-based technique can be applied to examples from RoboCup (autonomous robots play soccer). In experiments, it is shown how four-legged and humanoid robots can localize themselves very precisely on a soccer field. The localization is based on bearings to objects obtained from digital images. This thesis presents a new technique to recognize field lines which doesn''t need any pre-run calibration and also works when the field lines are partly concealed and affected by shadows.

An abstract knowledge representation of cognitive robots - as used for reasoning and planning - typically relies on symbols denoting objects of the world and states of affairs. The process of creating and maintaining the correct connection between a symbol denoting an object and its corresponding perceptual image (called percept), both referring to the same physical object, is called symbol anchoring. Most current cognitive systems implement an ad hoc solution which may work for the specific, intended application under certain conditions. Conversely, we suggest a formal and general approach to the symbol anchoring problem, which enhances previous approaches in terms of flexibility, applicability and expressiveness, and which completely automates the process of determining and maintaining all plausible hypotheses of correspondences between object symbols and perceptual images of physical objects. Based on the first-order logical Fluent Calculus, our approach inherits its rich expressiveness with respect to knowledge representation and reasoning. Implementing all required symbol anchoring functionalities, our approach also complies with fundamental concepts of phenomenalism, representationalism and the sense-data theory of philosophy of cognition.

Diese Arbeit befasst sich mit zwei ebenso alten wie bis in die Gegenwart bedeutsamen kontroversen Positionen in der Philosophie- und Wissenschaftsgeschichte: Einerseits dem Nativismus, der davon ausgeht, dass (wesentliche) Ideen in den Menschen eingeboren sind, hier vertreten durch die Position Noam Chomskys, dessen Arbeiten über universelle Grammatik die Linguistik im 20. Jahrhundert revolutionierten und weitreichende Auswirkungen auf Philosophie und Kognitionswissenschaften hatten. Dem gegenübergestellt wird die Position von Jean Piaget, der als Entwicklungspsychologe aus seiner Arbeit mit Kindern ein radikal konstruktivistisches Stufenmodell der Erkenntnisentwicklung vom Baby bis zum wissenschaftsfähigen Erwachsenen entwickelt hat, das auf angeborene Ideen verzichtet. Dargestellt werden historische und erkenntnistheoretische Hintergründe dieser interdisziplinären Streitfrage und die Argumente beider Positionen.