کنفرانس مهندسی برق ایران

صفحه اصلی / سی و سومین کنفرانس بین المللی مهندسی برق

Better Exploration In Single-Agent Q-Learning Using Controlled Linear Perturbation

نویسندگان :

Sadredin Hokmi¹ Mohammad Haeri²

1- Sharif university of technology 2- Sharif university of technology

کلمات کلیدی :

Q-learning،Exploration،Controlled Linear perturbation،Convergence rate،Maze،Cart-Pole

چکیده :

Reinforcement learning algorithms, especially model-free algorithms like Q-learning, have shown reliable results in finding optimal solutions for many real-time applications. However, challenges such as exploration in real-time and the convergence rate need to be addressed, and many researches have proposed algorithms to tackle these challenges. Algorithms like speedy Q-learning, Zap Q-learning, algorithms based on adding a regularization term, noise injection, and many others have been introduced. In this paper, an algorithm based on controlled linear perturbation is presented, which, according to the numerical results, can significantly reduce unnecessary explorations that are risky in real-time. Additionally, the proposed algorithm does not depend on the learning rate \mathbit{\alpha}, \mathbit{\gamma}, or changes in coefficients. However, to be effective, the parameters of the algorithm should be chosen within the correct range. The results of applying the proposed algorithm have been compared with three reliable algorithms: standard Q-learning, speedy Q-learning, and noise injection. These comparisons were conducted in a 9x9 maze scenario and in the cart-pole environment.