| Abstract [eng] |
Reinforcement learning (RL) is a key field between the artificial intelligence and learning theory, offering substantial potential to improve decision-making capabilities in a vast array of contexts. The continuous need for increased efficiency and efficacy in RL presents an ongoing and worthwhile challenge that can be impactful across numerous applications and domains, ranging from game theory and robotics to complex optimization problems and beyond. This work puts emphasis on understanding and enhancing the RL landscape. In-depth study of various RL methods and the dynamics between them reveals that many algorithms, are successors of previous ones, simply by combining new or existing features. This thesis is based on the idea of improving RL performance through new modifications of existing algorithms. The works starts by presenting an overview of key RL concepts and algorithms, followed by the proposition of potential improvements to original methods. A testing framework has been defined, with the purpose of rigorously evaluating the alterations and providing accurate and unbiased data whether these suggested modifications yield positive results. Multiple proposed modifications, such as temporal-position based sampling, heat-based sampling system and others were implemented. Finally, all of modifications were put to test across different OpenAI Gym environments based on the previously defined testing framework. The results of testing were carefully analyzed, providing interesting insights into correlation between proposed modifications and the results it yields. Finally, the evidence gathered throughout the research suggested that the hypothesized modifications were indeed providing superior performance under certain conditions. This not only underscores the potential of the proposed modifications, but also paves the way for future studies in RL enhancements. |
