Reinforcement learning to drive a car by pattern matching

In the paper the actual state of a system is presented that is aimed to learn driving autonomously different vehicles on different courses exclusively by visual input. The subsystem intelligent image processing allows, as required, to locate the road mark edges of each single image. A trained search algorithm allows optimal search speed and high recognition rate and consequently efficiently converts road mark edges into abstract complete situation descriptions (ACSDs) capturing in a storage limited way the current situation the vehicle is in. The subsystem pattern matching successfully retrieves similar situations to the current one based on a pattern matching algorithm. The pattern matching algorithm used in this subsystem is optimised for search speed on one hand and usage for road situations on the other hand. The subsystem reinforcement learning is still under implementation. However, a simple approach implemented so far allows already autonomous driving on a learning-by-knowledge-transfer basis promising further positive results in the area of autonomous driving based on pattern matching.