Quantifying Uncertainty in Complex Reinforcement Learning Scenarios

Addressing new challenges in reinforcement learning (RL) research requires identifying the most suitable algorithms, which involves their development and evaluation using various benchmarks. This paper presents a comparative analysis of two methodologies for classifying the complexity of RL problems, including real-world benchmarks and the existence of optimal policies. The exploration extends to the existence of optimal policies highlighting the significance of assumptions and methodologies in different studies. Additionally, two theories are presented that demonstrate conditions under which an optimal policy does not exist. A complexity classification based on these theories is introduced.