Samsung Electronics Co., Ltd.

Yi Tian Xu

Xue Liu

Di Wu

Gregory Lewis Dudek

Xi Chen

Intaik Park

Taeseop Lee

Hang Li

Jikun Kang

Ju Wang

Michael Jenkin

Seowoo Jang

Chengming Hu

Tianyu Li

Dmitriy Rivkin

Jimmy Li

The present disclosure provides for methods, apparatuses, and non-transitory computer-readable storage media for load balancing traffic scenarios by a network device. In an embodiment, a method includes training a plurality of learning agents to load balance a respective plurality of traffic scenarios to obtain a plurality of control policies. The method further includes performing at least one clustering iteration. Each clustering iteration includes selecting a pair of control policies and merging the pair of control policies into a clustered control policy that replaces the pair of control policies. The method further includes determining to stop the performing of the at least one clustering iteration when a quantity of control policies remaining in the plurality of control policies meets a predetermined value. The method further includes deploying to each base station of a plurality of base stations a corresponding control policy from the plurality of control policies.

Traffic scenario clustering and load balancing with distilled reinforcement learning policies

A method of forecasting a future load may include: obtaining source data sets and a target data set that have been collected from a plurality of source base stations and a target base station, respectively; among a plurality of source machine learning models, selecting at least one machine learn source model that has a traffic load prediction performance higher than that of a target machine learning model through a negative transfer analysis; obtaining model weights to be applied to the target machine learning model and the selected at least one source machine learning model via an attention neural network that is jointly trained with the target machine learning model and the selected source machine learning models; obtaining a load forecasting model for the target base station by combining the target machine learning model and the selected at least one source machine learning model according to the model weights; and predicting a future communication traffic load of the target base station based on the load forecasting model.

Method of load forecasting via attentive knowledge transfer, and an apparatus for the same

Rapid and data-efficient training of an artificial intelligence (AI) algorithm are disclosed. Ground truth data are not available and a policy must be learned based on limited interactions with a system. A policy bank is used to explore different policies on a target system with shallow probing. A target policy is chosen by comparing a good policy from the shallow probing with a base target policy which has evolved over other learning experiences. The target policy then interacts with the target system and a replay buffer is built up. The base target policy is then updated using gradients found with respect to the transition experience stored in the replay buffer. The base target policy is quickly learned and is robust for application to new, unseen, systems.

Load balancing using data-efficient learning

Transfer learning based on prediction determines a similarity between a source base station and a target base station. Importance of parameters is determined and training is adjusted to respect the importance of parameters. A lack of historical data is compensated by selecting a base station as source base station which has a larger amount of historical data.

Transfer learning of network traffic prediction model among cellular base stations

A method is provided. The method includes receiving a first dimension set, extracting a first latent feature set from the first dimension set, training a first base predictor based on the first feature set, generating a second dimension set based on the first dimension set, the second dimension set having fewer dimensions than the first dimension set, extracting a second latent feature set from the second dimension set, training a second base predictor based on the second feature set, and generating a traffic prediction based on the first base predictor and the second base predictor.

Communication load forecasting accuracy with adaptive feature boosting

Hybrid use of dual policies is provided to improve a communication system. In a multiple access scenario, when an inactive user equipment (UE) transitions to an active state, it may be become a burden to a radio cell on which it was previously camping. In some embodiments, hybrid load balancing is provided using a hierarchical machine learning paradigm based on reinforcement learning in which an LSTM generates a goal for one policy influencing cell reselection so that another policy influencing handover over active UEs can be assisted. The communication system as influenced by the policies is modeled as a Markov decision process (MDP). The policies controlling the active UEs and inactive UEs are coupled, and measureable system characteristics are improved. In some embodiments, policy actions depend at least in part on energy saving.

Hierarchical policy learning for hybrid communication load balancing

A method, computer program, and computer system are provided for load forecasting. Datasets corresponding to source machine learning models and a target domain base model are identified. A set of forecasting models corresponding to the identified datasets are learned. An ensemble model is determined from the learned set of forecasting models based on gradient boosting. An available resource is allocated based on the ensemble model.

Short-term load forecasting

An apparatus distributing communication load over a plurality of communication cells may select action centers from random cell reselection values, based on a standard deviation of an internet protocol (IP) throughout over the plurality of communication cells; input a first vector indicating a communication state of a communication system and a second vector indicating the standard deviation of the IP throughout of the plurality of communication cells, to a neural network to output a sum of the action centers and offsets as cell reselection parameters; and transmit the cell reselection parameters to the communication system to enable a base station of the communication system to perform a cell reselection based on the cell reselection parameters.

Growing community of inventors

Montreal, Canada

Yi Tian Xu