This thesis is structured in two main parts. In the first part, a taxonomy of reinforcement learning control strategies for building energy management is proposed based on previous literature. The objective is to provide a clear classification framework for the analyzed algorithms and to better understand their suitability for different applications. In the second part, a comparative analysis is conducted between a single-agent Proxi- mal Policy Optimization (PPO) algorithm and three multi–agent PPO-based approaches within the CityLearn environment. The goal is to evaluate their performance in terms of energy cost and carbon emissions at the single-building level, to analyze how load profiles change compared to a baseline and a Rule–Based Control (RBC), and to assess district-level Key Performance Indicators (KPIs) from a grid perspective.