关键词: 云模型, 最大均值差异, 指标迁移学习, 跨平台跨领域仿真评估, 小数据仿真评估, 指标聚合

Abstract:

In response to the problem of rare data samples in application experiment scenarios, this paper proposes an indicator transfer learning method based on cloud models and Maximum Mean Discrepancy (MMD), which transfers the indicator calculation model from typical simulation experiment scenarios to application experiment scenarios to meet the needs across platform and domain simulation evaluation. Using the maximum mean difference method to align the indicator distribution in the typical simulation experiment scenario to the indicator distribution in the application experiment scenario, thereby achieves indicator transfer, and by using cloud models based on a small number of examples for modeling and sampling, improves the efficiency of indicator transfer learning modeling. The effectiveness of our method has been verified through several indicator model transfer learning experiments from typical simulation experiment scenario to several application experiment scenarios. The distributions of target domain by our method are closer than those by Generative Adversarial Network transfer learning to the distributions of source domain. The transfer learning peformance of perception, cognition, decision and action ability indicators improves averagely 36.62% by Wasserstein distance measure.

Key words: cloud model, maximum mean discrepancy(MMD), indicator transfer learning, simulation assessment across platform and domain, simulation assessment on small data, indicator aggregation