Method and devices for transfer learning for inductive tasks in radio access network
A framework for inductive tasks in RANs includes a source learner, a target learner and an inference entity that can be merged into existing or future RAN functions. The signaling mechanism between different framework entities may also be merged into existing or future network interfaces. The framework can be reused by multiple inductive tasks to perform a variety of tasks such as prediction of some network phenomena or to improve certain KPIs in the network.
1. A network operator device performing as a source learner in a radio access network, RAN, the network operator device comprising:
processing circuitry; and
a memory containing instructions executable by the processing circuitry, the network operator device operative to:
prepare source inductive parameters, SIPs; and
supply the SIPs to a target learner in the RAN, wherein
the SIPs include an inductive model of an inductive task,
the SIPs enable the target learner to generate a production model of a RAN inductive task using the inductive model of the inductive task,
the inductive task and the RAN inductive task are different,
the target learner is configured to provide the production model to an inference entity in the RAN, and
the target learner and the inference entity are different logical entities.
2. The network operator device of claim 1 , wherein the inductive model is trained using first data and the production model is trained using second data which is different from the first data, and the first data and the second data is any one or more of:
the first data is simulated network data, and the second data is real network data;
the first data is a high volume of correlated low-cost data, and the second data is a low volume of more precise and expensive data than the high volume of correlated low-cost data;
the first data is offline available network data and the second data is online network data; and
the first data is 4G data, the inductive model being related to a 5G inductive task, and the second data is 5G data.
3. The network operator device of claim 1 , wherein the network operator device is further configured to train models included in SIPs and/or validate SIPs using data from the RAN.
4. The network operator device of claim 1 , wherein the RAN inductive task is a machine learning application predicting a certain network metric or measurement, classifying a network-related phenomenon or detecting an event or network status.
5. The network operator device of claim 1 , wherein the source learner is part of a data center of the RAN.
6. A network operator device performing as a source learner in a radio access network, RAN, the network operator device comprising:
processing circuitry; and
a memory containing instructions executable by the processing circuitry, the network operator device operative to:
prepare source inductive parameters, SIPs; and
supply the source inductive parameters, SIPs to a target learner in the RAN, wherein
the SIPs include an inductive model of an inductive task for a cell of a first cell type,
the SIPs enable the target learner to generate a production model of a RAN inductive task using the inductive model of the inductive task,
the inductive task and the RAN inductive task are different,
the target learner is configured to provide the production model to an inference entity in the RAN, and
the target learner and the inference entity are different logical entities.
7. The network operator device of claim 6 , wherein the RAN inductive task is for a cell of a second cell type.
8. The network operator device of claim 6 , wherein the inductive model is generated from data related to multiple cells and the production model is for a specific cell.