Reinforcement Learning based on visual information

OData support
Engedy István Tamás
Department of Measurement and Information Systems

Machine learning, including reinforcement learning plays an important role in our everyday life, sometimes without us even noticing it. Soon, our homes can fully manage themselves, our cars can drive to their destinations without any human interaction and most simple human tasks will be done by autonomous robots. AI algorithms do fraud detection and risk analysis, can collect the most important news for us based on our preferences and interests, make recommendations for us when browsing Netflix or Amazon, and can monitor our property or systems when we cannot. The list is far too long to be fully introduced here, but these few examples show that machine learning is currently a very hot topic, and it is in the focus of many researchers and market participants.

Besides reinforcement learning and machine learning, the field of visual data processing is going through a significant development as well. The reason behind this development is the fact that we need it. It is necessary to build machines that can check the quality of freshly manufactured stainless steel parts with a camera above the production line in the blink of an eye, or to run algorithms which can check and sort immense amounts of handwritten text, or to create systems which can save lives by helping doctors to make decisions by analysing radiograms.

In this thesis I created a small framework for combining reinforcement learning and visual information processing, and for testing different reinforcement learning algorithms on a simple game. During the running of these algorithms, the RL (Reinforcement Learning) agent learns by receiving only visual information about the game. I evaluated the results of these algorithms highlighting the differences between their efficiency in exploration.


Please sign in to download the files of this thesis.