CENG 715 : Evolutionary Computation
|
|
Context-based Reactive Controller with Evolved Probabilities |
The aim of this project is to evolve simple controllers which selects actions based on the current situations in a probabilistic manner, in the problem of self-aggregation in a swarm of robots. Self-aggregation is an important emergent behavior because it is the precursor of all other self-formations. Since robots have no direct communication, and their sensor modalities are too restricted, that they can perceive only their local, this problem has no ad-hoc, simple solution. Additionally, resulting system should behave robustly when any robot is broke down, and shoud be scalable to various size environments. |
CENG 710 : Introduction to Autonomous Robotics
|
|
Design of Reactive Behaviors for Unknown Environments using Potential Fields Methodology |
The aim of this work is the design and implementation of a set of behaviors for an autonomous, mobile robot to accomplish a given task. The task is to navigate towards a given target while avoiding from obstacles. To simulate the world, a simple 2D simulator is employed. Behaviors are designed in an Object-Oriented approach. Potential fields methodology is proved to work well in unstructured environments and reactive paradigm is successfully applied to solve the problem in the unknown world. |
|
|
Design of a Hybrid Control Architecture for Indoor Mobile Robot Navigation |
The aim of this project is the study of hybrid method in a sample navigation task. The task of the robot is to navigate in an indoor world and reach to a target point. The map of the world is known a priori, however there exists randomly distributed small obstacles. Robot plans, decomposes and executes its actions, following an optimal route, while avoiding from obstacles in a reflexive manner. A simple 2-D simulator is used in experiments. |
|
|
Particle Swarm Optimization (PSO) and PSO with Cross-over |
In this work, Particle Swarm Optimization (PSO) method is implemented and applied to various mathematical functional optimization and engineering problems. Then, we extended the original PSO method, and added the cross-over operator of the evolutionary computation techniques. Instead of applying the cross-over operator in each iteration, we evolved completely different swarms using PSO, and then later cross-overed particles from these different swarms. As a result, we obtained much better results with PSO-Cross, especially when the problem size is increased in the pure mathematical optimization problems. As an engineering problem, the welded beam design problem is also studied. |
|
|
A Learning Based Method for Super-Resolution of Low Resolution Images |
The main objective of this project is the study of a learning based method for super-resolving low resolution images. The domain specific prior is incorporated into super resolution by the means of learning based estimation of missing details. Images are decomposed into fixed size patches in order to deal with time and space complexity. The problem is modeled by Markov Random Field which enforces resulting images to be spatially consistent. The spatial interactions are coupled with a similarity constraint which should be established between high-resolution training image patches and low resolution observations. |
CENG 583 : Computational Vision
|
|
Shape From Shading: Deterministic and Stochastic Optimization |
The objective of this project is the study of Shape from Shading (SfS) problem in a variational framework. Among the other shape recovery methods, SfS is employed to obtain three-dimensional shape using spatial variations of brightness in an image. In this study, SfS problem will be addressed as an energy minimization problem. A constraint which establishes a connection between intensity gradient space and solution surface is imposed into the corresponding energy functional for the first time in a direct energy minimization scheme. A hybrid approach which combines the efficiency of deterministic methods and accuracy of stochastic methods is proposed in a multiresolution domain for the solution of this problem. |
|
|
A Review of Standard Back-Propagation Algorithm |
The aim of this work is to re-study the basic structure of the back-propagation learning algorithm. The other objective is to provide a complete description of this training algorithm with all the technical details. The algorithm is applied to two different benchmark problems and the network is analysed in various aspects. The experiments are mainly conducted to optimize the parameters of the system and to understand the relation between the initial parameter set and nature of the problem. |
|
|
Understanding Dynamics of Kohonen's Self-Organizing Feature Maps |
The aim of this work is to study the basics of an unsupervised learning method, Kohonen's algorithm, which works on top of a neural network whose units are geometrically related and activated in a winner-take-all basis. The network is capable of discovering and mapping the topological features in input patterns only using internal dynamics. Kohonen's Self-Organizing Map (SOM) is trained by a number of data sets which differ in their distribution on input space. One of the main objectives of this study is to optimize the parameters of the system while other objective is to analyze spatial ordering of output neurons with different data sets. |
|
|
Feed-Forward Competitive Shunting Nets: Visualization of Analytical Results |
The main objective of this work is to study the characteristics of a biologically plausible neural network model, feed-forward competitive shunting networks. Basic properties of this network is extracted from shunting equations and analytical results are visualized by employing some test cases in experiments. |
CENG 490 : Senior Design Seminar and Project
|
|
Evolution of probabilistic control in a swarm of robots |
The objective of this project is to study of a novel approach, design and implementation of a probabilistic architecture which is evolved to control simple behaviors in a swarm of robots. A simple feed-forward neural network(NN) is designed to control actuators or signalling units. Artificial evolutionary techniques, genetic algorithms or evolutionary strategies are used to train the network. The controller of the swarm is tested in a simple and fast robotic simulator. The practical outcome of the project is the emergence of robust and big cluster of robots. |