.png){kind=icon}
.png){kind=icon}
Return
MTPPO-PF: multi-task proximal policy optimization with particle filters for efficient hyperparameter optimization
.png){kind=icon}
PRE
AI
DOI：10.1007/s12293-025-00482-2.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Hyperparameter optimization (HPO) plays a crucial role in boosting the performance of machine learning models. Well-tuned hyperparameters can significantly enhance overall model performance. However, the complexity and nonlinear nature of the hyperparameter space makes it challenging for traditional optimization methods to capture interdependencies among parameters and achieve globally optimal solutions. To address this challenge, multi-task proximal policy optimization with particle filtering (MTPPO-PF) is introduced, which integrates a multi-task learning framework with proximal policy optimization to effectively model both the intrinsic dependencies and relative independence among hyperparameters. Additionally, particle filtering (PF) is employed to dynamically estimate task weights, enabling the optimizer to adaptively assess the relative importance of each hyperparameter in real time. Experimental results demonstrate that MTPPO-PF significantly outperforms existing mainstream methods across multiple tasks, offering notable advantages in optimization efficiency, accuracy, and robustness. Consequently, the approach provides an effective solution for addressing complex HPO challenges.
Keywords:
Hyperparameter optimization
Reinforcement learning
Particle filtering
Machine learning
Journal
IF:
2.3
Papers:
447
Citations:
718