Reinforcement learning is a branch of machine learning. It is dedicated to solve sequential decision problems in environments with unknown dynamics. During the learning process, it develops an optimal strategy in order to maximize some notion of cumulative reward. One of the central problems of reinforcement learning is the way to find balance between exploration and exploitation of the unknown environment.
In recent years, deep neural networks that are used in machine learning gained widespread popularity. Accordingly, new agents appeared besides the traditional methods, and they are able to learn using unprocessed images instead of preprocessed inputs, which led to a drastic increase of the state space.
Proper exploration algorithms are needed for the agent in order to learn efficiently. The algorithms that can be found in academic literature often do not perform well enough when problems present big state spaces, therefore other solutions were provided that are newer and can be more efficiently implemented.
This thesis gives an overview on the academic literature concerning this subject, introduces the used online environments, and demonstrates various newer exploration algorithms, then compares them to traditional, simpler, but widely used exploration algorithms. Afterwards, with the application of a deep neural network, it evaluates the implemented strategies according to performance and running time in three different environments.