New Criterion for Mapping Between Reinforcement Learning Frameworks

Different methodologies have been established by researchers to formalize reinforcement learning (RL). For an agent operating in one RL framework to function in another, it must undergo a conversion or mapping process. In a recent paper, the authors establish a foundation for examining the concept of relative-intelligence-preserving mappability across RL frameworks. They propose a criterion deemed sufficient for maintaining relative intelligence based on a specific intelligence measurement approach. Their findings indicate that this criterion is unattainable when transitioning between certain deterministic and stochastic RL frameworks, highlighting fundamental distinctions between these RL variations. The paper can be accessed on arXiv with the identifier 2112.07752 in the Computer Science > Artificial Intelligence section, including submission history and references from Semantic Scholar and DBLP. Additionally, it discusses arXivLabs, a platform for collaborative feature development, reinforcing arXiv's dedication to openness, community, excellence, and user data privacy.