This study investigates the application of Deep Reinforcement Learning (DRL) techniques in the Arcade Learning Environment (ALE), with the aim of developing agents capable of outperforming humans in Atari 2600 games. The research focuses on evaluating the performance and convergence of wellestablished weight and bias initialization techniques in the literature within the architecture of the Deep Q-Network (DQN). The analysis includes comparisons between different initialization strategies and their implications on the learning efficiency and robustness of agents trained across a range of games.