Contributions published at Artificial Intelligence Lab

This paper will give some insight into a program that was developed for the purpose of tracking a moving object. A picture, showing the object that is to be tracked, is the point of departure. The SIFT-algorithm uses this reference image to find the object from images that were extracted from a movie. The result of this step are coordinates of potential matches, where the object being tracked is on the picture. This data is then used to calculate the mean coordinates of the matches. These values are used to calculate the object’s velocity and the distance that the robot has moved. This paper puts special emphasis on the problems that have to be tackled in order to produce a program that works and how these problems were solved.

The goal of this thesis is to generate algorithms to forecast the visibility in meters at Zurich airport. This visibility indicates the intensity of fog. The new implemented approaches to predict fog are genetic programming and evolutionary strategies. A MOS (model output statistic) is computed and evaluated. The MOS results are used as an input variable for genetic programming. The best solution is compared with the forecasts of the meteorologists in four categories: Proportion correct, probability of detection, false alarm ratio and the Heidke skill score. In all four categories the performance of the meteorologists are a little better.

In robotics the adaptability to dynamic environments and the ability for on the spot repairs is a key challenge for engineers. Modular robotics has, inspired by living organisms and natural processes, tried to overcome this problem by using self-repair and self-assembly to achieve the task through local interaction and control. The main contribution of this work is the mathematical analysis of a stochastically self-assembling system (Tribolon) which was developed at the AI-Lab; University of Zurich .The platform features cm-sized modules floating on water. It allows the study of aggregation behavior of a large number of embodied components in a physical environment. In our evaluation we employ kinetic rate equations to analyze the dynamics of the system with a large number of modules. We observe that the system converges to an intermediate state where not all clusters have reached their maximally possible size, i.e., forming incomplete circles. This is known as the yield problem. We further examine the influence of the modules’ angle on the rate of fully formed circles. Our calculations point to a power-law relationship between angle and yield rate. We also introduce the concept of degree of parallelism (DOP), to study the connections within a cluster of modules. We conclude that clusters following an aggregation pathway with higher DOP will be less likely to reach a full circle configuration. Finally we have used the finite element method to investigate the magnetic properties of our clusters. We show that by varying the positioning of magnets in our modules we can influence cluster stability. In conclusion our results hint at a novel way to control self-assembly processes. By varying morphological features such as the angle of our modules and positioning of the magnet we can influence the final yield rate of a process. We therefore propose that morphology should also be considered as a control mechanism in stochastic modular robotics.

This work analyses the potential use of genetic programming (GP) in the storm warning system of MeteoSwiss. The chosen test object is Grenchen, an airport in Switzerland with clear meteorological characteristics. Different settings are tested in order to find an optimized GP algorithm for the described problem domain. The customer costs are an important factor in order to produce customer cost optimizing warnings. Therefore they are fed into the GP algorithm. A new fitness measure based on performance figures for warning systems is introduced. Dimensionally aware GP is implemented and shown to be able to produce better solutions than traditional GP in this physics related problem domain. A Genetic Ensemble Warning System (GEWS) is introduced in order to take several GP runs into account for the warning decision. This approach stands in contrary to traditional GP which only uses one solution as base of decision. This thesis shows that a GEWS produces better performing solutions than traditional GP. The ability of GP to generate storm warnings in a human competitive way is clearly demonstrated.

It is widely agreed that amputees have to rely on visual input to monitor and control the position of the prosthesis while reaching and grasping because of the lack of proprioceptive feedback. Therefore, visual information has been a prerequisite for prosthetic hand biofeedback studies. This is why, the underlying characteristics of other artificial feedback methods used to this day, such as auditive, electro-tactile, or vibro-tactile feedback, has not been clearly explored. The purpose of this paper is to explore whether it is possible to use audio feedback alone to convey more than one independent variable (multichannel) simultaneously, without relying on the vision, to improve the learning of a new perceptions, in this case, to learn and understand the artificial proprioception of a prosthetic hand while reaching. Experiments are conducted to determine whether the audio signals could be used as a multi-variable dynamical sensory substitution in reaching movements without relying on the visual input. Two different groups are tested, the first one uses only audio information and the second one uses only visual information to convey computer-simulated trajectories of two fingers.

How is our body imprinted in our brain? This seemingly simple question is a subject of investigations of diverse disciplines, psychology, and philosophy originally complemented by neurosciences more recently. Despite substantial efforts, the mysteries of body representations are far from uncovered. The most widely used notions-body image and body schema-are still waiting to be clearly defined. The mechanisms that underlie body representations are coresponsible for the admiring capabilities that humans or many mammals can display: combining information from multiple sensory modalities, controlling their complex bodies, adapting to growth, failures, or using tools. These features are also desirable in robots. This paper surveys the body representations in biology from a functional or computational perspective to set ground for a review of the concept of body schema in robotics. First, we examine application-oriented research: how a robot can improve its capabilities by being able to automatically synthesize, extend, or adapt a model of its body. Second, we summarize the research area in which robots are used as tools to verify hypotheses on the mechanisms underlying biological body representations. We identify trends in these research areas and propose future research directions.

null

Self-Assembly is a process where simple parts autonomously form organized structures, simulations are used in research to find or optimize new promising parts and structures. This thesis describes the extension of a self-assembly simulator by a graphical user interface to configure experiments and the reengineering of certain parts of the simulator. The development is documented in three parts: requirements, design and implementation. In the end an evaluation of the chosen solution is made and an outlook is given on how to further improve the application.

This document describes in detail the developed module for generating evaluations of personal or team feedbacks. These feedbacks are being generated by the main webobject web application, which provides an xml file with all required data to the module. The module should mainly replace the currently implemented method of generating evaluations and bring some new features. Apart from that the windows server, which is being used for the generation of evaluations, should not be used any more. The generation will take place only in the main server. The module receives a xml file from the main application and generates several html pages. The customer can directly view them online. It will also create a PDF version of the evaluation. Finally also a zip package with all needed resources for the offline view of the evaluation will be generated. Design changes are very easy to manage, as only some small modifications in the CSS file or html code are required.

Living Technology is researching novel IT making strong use of programmable chemical systems. These chemical systems shall finally converge to artificial cells resulting in evolvable complex information systems. We focus on procedural manageability and information processing capabilities of such information systems. Here, we present a novel resource-saving formation, processing, and examination procedure to generate and handle single compartments representing preliminary stages of artificial cells. Its potential is exemplified by testing the influence of different glycerophospholipids on the stability of the compartments. We discuss how the procedure could be used both in evolutionary optimization of self-assembling amphiphilic systems and in engineering tailored communication networks enabling life-like information processing in multicompartment aggregates of programmable composition and spatial configuration.

Embryogenic evolution emulates in silico cell-like entities to get more powerful methods for complex evolutionary tasks. As simulations have to abstract from the biological model, implicit information hidden in its physics is lost. Here, we propose to use cell-like entities as a real-world in vitro testbed. In analogy to evolutionary robotics, where solutions evolved in simulations may be tested in real-world on macroscale, the proposed vesicular testbed would do the same for the embryogenic evolutionary tasks on mesoscale. As a first step towards a vesicular testbed emulating growth, cell division, and cell differentiation, we present a modified vesicle production method, providing custom-tailored chemical cargo, and present a novel self-assembly procedure to provide vesicle aggregates of programmable composition.

This projects implements a bio-inspired computational mechanism that allows the bidpedal hopping robot Stumpy to move in reaction to music, i.e. to dance. For that we first realize a way to distinguish features in music, namely the rhythmical pattern observant in music pieces. We identify the beat as a way to generate rhythmic behavior and we then use the beats of a music track as a marker and input the energy values of those beats into a network of FitzHugh-Nagumo neurons. These neurons, for the most part, spike periodically according to their input spike trains. Yet there are values for the input impulse and the periodicity of the impulse that lead the neurons to spike chaotically. The motor neurons of the neural network will then periodically output values which are then used to generate a parameter set to control Stumpy. We implement this process as a EffectsPlugin in XMMS using C and C++ source files. The plugin initializes a neural network, grabs packages of samples from the playback, performs a beat detection, feeds the energy values of beats periodically to the beat rate and transmits control parameter sets to Stumpy through a bluetooth connection. In the end we have a Stumpy that moves to the music played, which can be described as a rudimentary form of dancing.

null

null

null

In the last decades great progess was made to explain the evolution of cooperation. Although important strides were made in explaining why natural selection can favor cooperators over defectors, still many questions remain unanswered. This work is about the evolution of cooperation and is based on an agent simulation proposed by Mikahil Burtsev & Peter Turchin to investigate the evolution of cooperative strategies among simple agents. The model comprises of simple agents that live on cells that are arranged as a 2-dimensional grid. The agents can perform certain actions like moving, eating, dividing and ghting. They are controlled by a simple neural network that connects the inputs, that provide the agent with information about the outside world and its internal state (resource detection, agent detection, internal energy and so on), to the actions. The neural network is subject to mutation whenever an agent divides into osprings. I made some extensions to this original model among others the introduction of a give action that allows agents to transfer energy between each other and the addition of a reputation value that indicates whether an agent is cooperative (giving) or defective (attacking). The results of the simulations of this extended model show that the system actually can converge to a state where the agents use strategies that frequently make use of the give action. Analysis of the interaction network reveals that the agents organize in clusters of cooperation, where they mutually exchange energy.

The ECCEROBOT project is an effort to build and control an anthropomimetic robot which contains a muscule-tendon like actuator system embedded in a skeleton. This diploma thesis has been written in the context of the ECCEROBOT project and its goal was to simulate the working of a muscle as accurately as possible. To build the simulation, the first step was to select a suitable simulation tool. A survey of the available physics based simulation tools was undertaken to select the most suitable open-source tool with respect to accuracy, stability and speed of calculation for simulations of the robot. Finally, JBullet, the Java version of the Bullet physics engine, was selected in combination with OpenGL to simulate an arm of the robot. The simulation of the arm muscle was done based on the Hill model. Various combinations of parameters were investigated for the control of the muscle and also what actions the muscle could perform most effectively in simulation. The report contains the findings related with the accuracy of the simulated arm and its performance in interaction with other objects.