Proseminar Künstliche Intelligenz
Artificial Intelligence is a central discipline nowadays in computer science. Research at the Institute of Artificial Intelligence focusses on knowledge-based techniques, i.e., how can we formalize and reason about knowledge within intelligent systems. Knowledge-based techniques complement learning-based techniques, e.g., to provide far-ranging foresight based on map knowedge for autonomous vehicles or for planning actions that an intelligent agent needs to perform in order to reach a certain goal. The seminar will introduce students to selected research topics in this area.
Because meaningful sentences are composed of meaningful words, any system that hopes to process natural languages as people do must have information about words and their meanings. This information is traditionally provided through dictionaries, and machine-readable dictionaries are now widely available. But dictionary entries evolved for the convenience of human readers, not for machines. WordNet provides a more effective combination of traditional lexicographic information and modern computing. WordNet is an online lexical database designed for use under program control. English nouns, verbs, adjectives, and adverbs are organized into sets of synonyms, each representinga lexicalized concept. Semantic relations link the synonym sets.
Betreuer: Michael Weldt
In classical planning, the task is to drive a system from a given initial state into a goal state by applying actions whose effects are deterministic and known. Classical planning can be formulated as search problem whose nodes represent the states of the system or enviroment, and whose edges capture the state transitions that the actions make possible. State-of-the-art methods in classical planning search towards the goal using heuristic functions that are automatically derived from the problem.
https://www.morganclaypool.com/doi/abs/10.2200/S00513ED1V01Y201306AIM022 (nur Teile von Kapitel 2)
Betreuer: Conny Olz
The Internet of Musical Things (IoMusT) is an emerging research area consisting of the extension of the Internet of Things paradigm to the music domain. Interoperability represents a central issue within this domain, where heterogeneous objects dedicated to the production and/or reception of musical content (Musical Things) are envisioned to communicate between each other. This paper proposes an ontology for the representation of the knowledge related to IoMusT ecosystems to facilitate interoperability between Musical Things. There was no previous comprehensive data model for the IoMusT domain, however the new ontology relates to existing ontologies, including the SOSA Ontology for the representation of sensors and actuators and the Music Ontology focusing on the production and consumption of music. This paper documents the design of the ontology and its evaluation with respect to specic requirements gathered from an extensive literature review, which was based on scenarios involving IoMusT stakeholders, such as performers and audience members.
Schwarmintelligenz beschreibt ein natürliches Phänomen, bei dem Individuen, trotz ihrer oftmals eingeschränkten Fähigkeiten, dennoch in der Lage sind, komplexe Aufgaben als vereintes Kollektiv zu lösen. Ein typisches Beispiel hierfür ist die Wegfindung von Ameisen, wobei aufgrund der indirekten Interaktion mittels Pheromonspuren, immer der kürzeste bzw. effektivste Weg zu einer Futterquelle gefunden wird. In der Informatik findet man Schwarmintelligenz hauptsächlich im Kontext von Optimierungsproblemen wieder, wofür Algorithmen, wie zum Beispiel Ant Colony Optimization (Dorigo) oder Particle Swarm Optimization (Eberhart, Kennedy, Shi), zum Einsatz kommen.
Ambient Intelligence beschäftigt sich damit, intelligente Technologien in Umgebungen zu integrieren, um besser auf die Bedürfnisse der Benutzer eingehen zu können. Hierfür werden, einerseits, mittels Sensoren, verschiedene Daten über die Umgebung und deren Entitäten gesammelt, während man, andererseits, Methoden aus der KI, wie z.B. Reasoning, verwendet, um diese Daten zu verarbeiten. Typische Beispiele sind Smart Homes (Einschalten der Heizung, wenn Benutzer von Arbeit kommt), e-Health (Notruf, wenn Benutzer sich verletzt) oder Smart Cities (Regulierung von Verkehr zur Verhinderung von Staus).
Methods for unsupervised hypernym detection may broadly be categorized according to two paradigms: pattern-based and distributional methods. In this paper, we study the performance of both approaches on several hypernymy tasks and find that simple pattern-based methods consistently outperform distributional methods on common benchmark datasets. Our results show that pattern-based models provide important contextual constraints which are not yet captured in distributional methods.
Betreuer: Michael Welt
Games have always been a fertile ground for advancements in computer science, operations research and artificial intelligence. Solitaire card games, and Freecell in particular, have been the subject of study in both the academic literature, where they are used as a benchmark for planning heuristics, and in popular literature. Here an approach which provides provably optimal solutions to solitaire games shall be studied. It uses A* search together with an admissible heuristic function that is based on analyzing a directed graph whose cycles represent deadlock situations in the game state.
Betreuer: Conny Olz
Solving games with AI has been an longstanding research benchmark in the field. While chess was “solved”, go has been elusive due to the much larger search space. In this work we want to take a look how large search-spaces can be searched efficiently with learned heuristics.
Traditionally for each different NLP Task, a new system had to be designed and trained which is cumbersome and often data-hungry. These problems limit the application in smaller domains. But what if one could use a single model and dataset for a variety of tasks and additionally transfer these models towards different domains?
A central problem in designing and implementing interactive systems---action selection---is also a core research topic in automated planning. While numerous toolkits are available for building end-to-end interactive systems, the tight coupling of representation, reasoning, and technical frameworks found in these toolkits often makes it difficult to compare or change the underlying domain models. In contrast, the automated planning community provides general-purpose representation languages and multiple planning engines that support these languages. We describe our recent work on automated planning for task-based social interaction, using a robot that must interact with multiple humans in a bartending domain.
Betreuer: Milene Santos Teixeira
Social robots working in public space often stimulate children’s curiosity. However, sometimes children also show abusive behavior toward robots. In our case studies, we observed in many cases that children persistently obstruct the robot’s activity. Some actually abused the robot by saying bad things, and at times even kicking or punching the robot. We developed a statistical model of occurrence of children’s abuse. Using this model together with a simulator of pedestrian behavior, we enabled the robot to predict the possibility of an abuse situation and escape before it happens. We demonstrated that with the model the robot successfully lowered the occurrence of abuse in a real shopping mall.
Betreuer: Felix Lindner
Despite widespread adoption, machine learning models remain mostly black boxes. Understanding the reasons behind predictions is, however, quite important in assessing trust, which is fundamental if one plans to take action based on a prediction, or when choosing whether to deploy a new model. Such understanding also provides insights into the model, which can be used to transform an untrustworthy model or prediction into a trustworthy one. In this work, we propose LIME, a novel explanation technique that explains the predictions of any classifier in an interpretable and faithful manner, by learning an interpretable model locally around the prediction. We also propose a method to explain models by presenting representative individual predictions and their explanations in a non-redundant way, framing the task as a submodular optimization problem. We demonstrate the flexibility of these methods by explaining different models for text (e.g. random forests) and image classification (e.g. neural networks). We show the utility of explanations via novel experiments, both simulated and with human subjects, on various scenarios that require trust: deciding if one should trust a prediction, choosing between models, improving an untrustworthy classifier, and identifying why a classifier should not be trusted.