Analysis in growing data-driven fashions for Air Site visitors Administration (ATM)
has gained an amazing curiosity in recent times. Nonetheless, data-driven fashions
are identified to have lengthy coaching time and require massive datasets to attain good
efficiency. To handle the 2 points, this paper proposes a Multi-Agent
Bidirectional Encoder Representations from Transformers (MA-BERT) mannequin that
absolutely considers the multi-agent attribute of the ATM system and learns air
site visitors controllers’ selections, and a pre-training and fine-tuning switch
studying framework. By pre-training the MA-BERT on a big dataset from a serious
airport after which fine-tuning it to different airports and particular air site visitors
functions, a considerable amount of the overall coaching time will be saved. In
addition, for newly adopted procedures and constructed airports the place no
historic knowledge is on the market, this paper exhibits that the pre-trained MA-BERT can
obtain excessive efficiency by updating usually with little knowledge. The proposed
switch studying framework and MA-BERT are examined with the automated dependent
surveillance-broadcast knowledge recorded in 3 airports in South Korea in 2019.