Generalization performance improvement of artificial intelligence deep reinforcement learning algorithms for large-scale high-dimensional decision making
Year: 2026
Type: article
Abstract: In large-scale high-dimensional decision-making scenarios, the system needs to maintain the policy output under the expansion of state, the intensive combination of actions and the change of task distribution. To improve the generalization performance of deep reinforcement learning, a generalization enhancement method driven by state-action collaboration and task distribution training was proposed. Three types of tasks, industrial scheduling, resource allocation and continuous control, were set up in the experiment, and 18 groups of training tasks and 12 groups of unseen test tasks were constructed. The state dimension covered 128, 256 and 512 dimensions, and the action scale could reach ten thousand levels. The results show that the average reward of unseen tasks reaches 0.858, the success rate reaches 92.8%, the generalization gap is reduced to 11.6%, the constraint violation rate is reduced to 3.4%, and the inference delay is 24.1ms. Ablation experiments show that the success rate decreases by 5.6% points after removing state-action collaboration, and the generalization gap increases by 7.1% points after removing task distribution training. The results show that the proposed method improves the transfer generalization stability of the model in high-dimensional decision-making.
Source: Discover Applied Sciences
Author: Hanyu Zhang
Institution: South China University of Technology
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7226983184
