Track Learning Agent Using Multi-objective Reinforcement Learning

Reinforcement learning (RL) enables agents to make decisions through interactions with their environment and feedback in the form of rewards or penalties. The distinction between single-objective reinforcement learning (SORL) and multi-objective reinforcement learning (MORL) is established, emphasizing the latter's ability to optimize multiple conflicting objectives simultaneously. The study explores various algorithms and approaches within the MORL framework, focusing on track navigation optimization. Key components of the implementation are detailed, including states, actions, rewards, and tracks used for training. The proposed algorithm, Pareto Q-learning, is highlighted as a powerful approach to simultaneously optimize multiple objectives. The architecture and methodology of the learning agent are presented, outlining the training process and the impact of hyper-parameters. Results from experimentation are discussed, revealing the agent's learning curve, crash avoidance, and successful achievement of multiple objectives. The study underlines the significance of MORL in enabling agents to manage complex decision-making scenarios, leading to more robust and optimal policies. The paper concludes by emphasizing the practical implications of the MORL approach in navigating challenging tracks with conflicting goals, such as minimizing steps, maximizing rewards, and avoiding collisions.