Scholarlink

<p>Deep reinforcement learning is known as a revolutionary technique in the field of computer games and the design of intelligent systems in recent years. However, ensuring stability and robustness of learning agents in sparsely and irregularly rewarded environments has always been a complex and open issue, even with modern improvements. In this work, it is shown that in the Flappy Bird game environment using DQN, Double DQN, Dueling DQN learning algorithms with soft updating and frame stacking, the stability of the learning agent can be increased and the performance more accurate. It is proved that the parallel use of these algorithms and methods leads to lower fluctuations in performance metrics, higher training stability, and higher final agent performance. The results of this research can be used to design more stable and practical reinforcement learning systems in complex real-world environments. The proposed hybrid approach achieves a maximum score of approximately 180 in the Flappy Bird environment and demonstrates improved convergence behavior compared to baseline methods. Furthermore, the approach significantly reduces performance variance, resulting in more stable training behavior across different runs.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A hybrid approach to reinforcement learning: combining soft updating and frame stacking for stable training in Flappy Bird and CartPole

  • 期刊论文

  • 发表日期：

    2026年6月15日

查看全文
  • Mohammad Hossein Jabbarzadeh,
  • Delaram Mehralizadeh,
  • Kimia Shirini,
  • Sina Samadi Gharehveran

摘要

Deep reinforcement learning is known as a revolutionary technique in the field of computer games and the design of intelligent systems in recent years. However, ensuring stability and robustness of learning agents in sparsely and irregularly rewarded environments has always been a complex and open issue, even with modern improvements. In this work, it is shown that in the Flappy Bird game environment using DQN, Double DQN, Dueling DQN learning algorithms with soft updating and frame stacking, the stability of the learning agent can be increased and the performance more accurate. It is proved that the parallel use of these algorithms and methods leads to lower fluctuations in performance metrics, higher training stability, and higher final agent performance. The results of this research can be used to design more stable and practical reinforcement learning systems in complex real-world environments. The proposed hybrid approach achieves a maximum score of approximately 180 in the Flappy Bird environment and demonstrates improved convergence behavior compared to baseline methods. Furthermore, the approach significantly reduces performance variance, resulting in more stable training behavior across different runs.