Splitting and Parallel Training of Dense Layer in Dueling DQN

---

Project Team : Ankit Kumar (121ad0005@iiitk.ac.in) Rajiv Kumar (121ad0043@iiitk.ac.in)

Solving real-life problems faster and more accurately with AI & ML models benefits us in numerous ways. To enhance this, we are modifying the Dueling DQN structure by splitting the dense layer into multiple layers, creating separate advantage and value streams that converge at the final layer. This allows for parallel training: one stream estimates the state value V(s), while the other estimates the advantage for each action A(s, a).

This approach increases the convergence rate, leading to faster training and quicker performance improvements. It also boosts accuracy in problem-solving and enhances applications in robotics and gaming by improving action-taking decisions.

Back