Optimizing Reinforcement Learning Using Failure Data

Learning from both successes and failures is key to developing robust and efficient policies in reinforcement learning (RL). Traditional RL excels at learning from rewards but often neglects non-rewarding states, especially those leading to negative outcomes. This paper introduces a novel approach that integrates a modified Gaussian distribution into a Deep Q-Network (DQN) framework to learn from failures. By penalizing state-action pairs near historical failure points, the model guides the agent away from pitfalls. The optimized DQN shows improved learning speed and stability, achieving higher and more consistent scores than a standard DQN. This approach highlights the potential of hybrid RL models that combine value-based methods with failure-aware mechanisms to accelerate learning and enhance decision-making.