This thesis explores the application of reinforcement learning in training artificial intelligence agents to navigate a racing game environment. Using Unity's ML-Agents Toolkit and the Proximal Policy Optimization (PPO) algorithm, the study investigates how hyperparameter tuning, reward engineering, and parallel training impact agent performance. A simplified Mario Kart–inspired environment was used to simulate real-time driving challenges. By systematically modifying learning rates, batch sizes, and discount factors, and analyzing their effects, the research identifies optimal training configurations. Additionally, custom reward functions significantly influenced learning efficiency and agent behavior, highlighting the importance of reward design. The findings demonstrate how RL techniques can be effectively applied to both gaming and real-world autonomous driving scenarios.