Scholarlink

In this paper, in order to make the near-lunar space decision-making more reliable, a new neural network model is proposed that combines an Improved Bayesian Neural Network (IBNN) with a Multi-Task Unified Transformer (MT-UT) architecture. Based on the MT-UT’s task-adaptive mechanism, multi-task learning difficulties of interdependent tasks in near-lunar circumstances are solved. On the other hand, by making using of the Gaussian mixture priors to model weight distributions, a more precise measurement of the uncertainty in model predictions is enabled in the enhanced BNN. Experimental results show that the proposed MT-UT + IBNN model performs better than single-task deep learning and classical machine learning baselines across all assessment measures with an F1 score of 0.87 and a false positive rate of less than 8%, thereby improving the reliability of near-lunar decision-making.

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

Bayesian Optimization Based Multi-task Unified Transformer for Near-Lunar Space Decision-Making

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yu Lei,
  • Di Wang,
  • Siliang Yang

摘要

In this paper, in order to make the near-lunar space decision-making more reliable, a new neural network model is proposed that combines an Improved Bayesian Neural Network (IBNN) with a Multi-Task Unified Transformer (MT-UT) architecture. Based on the MT-UT’s task-adaptive mechanism, multi-task learning difficulties of interdependent tasks in near-lunar circumstances are solved. On the other hand, by making using of the Gaussian mixture priors to model weight distributions, a more precise measurement of the uncertainty in model predictions is enabled in the enhanced BNN. Experimental results show that the proposed MT-UT + IBNN model performs better than single-task deep learning and classical machine learning baselines across all assessment measures with an F1 score of 0.87 and a false positive rate of less than 8%, thereby improving the reliability of near-lunar decision-making.