초록 ▼

A server uses an LSTM neural network to predict a bandwidth value for a computer network element using past traffic data. The server receives a time series of bandwidth utilization of the computer network element. The time series includes bandwidth values associated with a respective time values. Th

A server uses an LSTM neural network to predict a bandwidth value for a computer network element using past traffic data. The server receives a time series of bandwidth utilization of the computer network element. The time series includes bandwidth values associated with a respective time values. The LSTM neural network is trained with a training set selected from at least a portion of the time series. The server generates a predicted bandwidth value associated with a future time value based on the LSTM neural network. The provisioned bandwidth for the computer network element is adjusted based on the predicted bandwidth value.

대표청구항 ▼

1. A computer-implemented method comprising: receiving a time series of bandwidth utilization of a computer network element, the time series comprising a plurality of bandwidth values each associated with a respective time value;training a Long Short Term Memory (LSTM) neural network with a training

1. A computer-implemented method comprising: receiving a time series of bandwidth utilization of a computer network element, the time series comprising a plurality of bandwidth values each associated with a respective time value;training a Long Short Term Memory (LSTM) neural network with a training set comprising at least a portion of the time series;generating a predicted bandwidth value associated with a future time value based on the LSTM neural network; andadjusting a provisioned bandwidth for the computer network element based on the predicted bandwidth value. 2. The method of claim 1, further comprising iteratively predicting a subsequent bandwidth value by: updating the training set with the predicted bandwidth value associated with the future time value;retraining the LSTM neural network with the updated training set; andgenerating the subsequent bandwidth value associated with a subsequent future time value based on the retrained LSTM neural network. 3. The method of claim 1, wherein the training set comprises raw bandwidth data without any decomposition. 4. The method of claim 1, wherein the LSTM neural network is trained with the training set without removing a periodic seasonality. 5. The method of claim 1, wherein the LSTM neural network comprises a plurality of blocks, each block comprising an input gate, a forget gate, a memory cell, and an output gate. 6. The method of claim 5, wherein a state of a particular block of the LSTM neural network at a time t given an input x(t) is defined by: i(t)=gi(x(t)·Wxi+h(t−1)·Whi+c(t−1)·Wci+bi)f(t)=gf(x(t)·Wxf+h(t−1)·Whf+c(t−1)·Wcf+bf)c(t)=f(t)·c(t−1)+i(t)·tan h(x(t)·Wxc+h(t−1)·Whc+bc)o(t)=go(x(t)·Wxo+h(t−1)·Who+c(t)·Wco+bo)h(t)=o(t)·tan h(c(t))y(t)=gy(h(t)·Why+by),where ga(A) is a sigmoid function, i(t) is an input gate state, f(t) is a forget gate state, c(t) is a memory cell state, o(t) is an output gate state, h(t) is a hidden layer output, y(t) is an output of the particular block, Wzq is a weight matrix that connects element z to element q, and bz is a bias term for element z. 7. The method of claim 6, wherein the sigmoid function ga(A) is defined by: ga(A)=Ja+Ka1+e-LaA,where Ja, Ka, and La are parameters determined by the training of the LSTM neural network. 8. An apparatus comprising: a network interface unit configured to communicate with computer network elements in a computer network;a Long Short Term Memory (LSTM) neural network configured to process a time series of bandwidth utilization received via the network interface unit, the time series comprising a plurality of bandwidth values each associated with a respective time value; anda processor configured to: train the LSTM neural network with a training set comprising at least a portion of the time series;generate a predicted bandwidth value associated with a future time value based on the LSTM neural network; andadjust a provisioned bandwidth for at least one of the computer network elements based on the predicted bandwidth value. 9. The apparatus of claim 8, wherein the processor is further configured to iteratively predict a subsequent bandwidth value by: updating the training set with the predicted bandwidth value associated with the future time value;retraining the LSTM neural network with the updated training set; andgenerating the subsequent bandwidth value associated with a subsequent future time value based on the retrained LSTM neural network. 10. The apparatus of claim 8, wherein the processor is configured to train the LSTM neural network with the training set without removing a periodic seasonality. 11. The apparatus of claim 8, wherein the training set comprises raw bandwidth data without any decomposition. 12. The apparatus of claim 8, wherein the LSTM neural network comprises a plurality of blocks, each block comprising an input gate, a forget gate, a memory cell, and an output gate. 13. The apparatus of claim 12, wherein a state of a particular block of the LSTM neural network at a time t given an input x(t) is defined by: i(t)=gi(x(t)·Wxi+h(t−1)·Whi+c(t−1)·Wci+bi)f(t)=gf(x(t)·Wxf+h(t−1)·Whf+c(t−1)·Wcf+bf)c(t)=f(t)·c(t−1)+i(t)·tan h(x(t)·Wxc+h(t−1)·Whc+bc)o(t)=go(x(t)·Wxo+h(t−1)·Who+c(t)·Wco+bo)h(t)=o(t)·tan h(c(t))y(t)=gy(h(t)·Why+by),where ga(A) is a sigmoid function, i(t) is an input gate state, f(t) is a forget gate state, c(t) is a memory cell state, o(t) is an output gate state, h(t) is a hidden layer output, y(t) is an output of the particular block, Wzq is a weight matrix that connects element z to element q, and bz is a bias term for element z. 14. The apparatus of claim 13, wherein the sigmoid function ga(A) is defined by: ga(A)=Ja+Ka1+e-LaA,where Ja, Ka, and La are parameters determined by the training of the LSTM neural network. 15. One or more non-transitory computer readable storage media encoded with computer executable instructions operable to cause a processor to: receive a time series of bandwidth utilization of a computer network element, the time series comprising a plurality of bandwidth values each associated with a respective time value;train a Long Short Term Memory (LSTM) neural network with a training set comprising at least a portion of the time series;generate a predicted bandwidth value associated with a future time value based on the LSTM neural network; andadjust a provisioned bandwidth for the computer network element based on the predicted bandwidth value. 16. The computer readable storage media of claim 15, further comprising instructions operable to cause the process to iteratively predict a subsequent bandwidth value by: updating the training set with the predicted bandwidth value associated with the future time value;retraining the LSTM neural network with the updated training set; andgenerating the subsequent bandwidth value associated with a subsequent future time value based on the retrained LSTM neural network. 17. The computer readable storage media of claim 15, wherein the training set comprises raw bandwidth data without any decomposition. 18. The computer readable storage media of claim 15, wherein the LSTM neural network comprises a plurality of blocks, each block comprising an input gate, a forget gate, a memory cell, and an output gate. 19. The computer readable storage media of claim 18, wherein a state of a particular block of the LSTM neural network at a time t given an input x(t) is defined by: i(t)=gi(x(t)·Wxi+h(t−1)·Whi+c(t−1)·Wci+bi)f(t)=gf(x(t)·Wxf+h(t−1)·Whf+c(t−1)·Wcf+bf)c(t)=f(t)·c(t−1)+i(t)·tan h(x(t)·Wxc+h(t−1)·Whc+bc)o(t)=go(x(t)·Wxo+h(t−1)·Who+c(t)·Wco+bo)h(t)=o(t)·tan h(c(t))y(t)=gy(h(t)·Why+by),where ga(A) is a sigmoid function, i(t) is an input gate state, f(t) is a forget gate state, c(t) is a memory cell state, o(t) is an output gate state, h(t) is a hidden layer output, y(t) is an output of the particular block, Wzq is a weight matrix that connects element z to element q, and bz is a bias term for element z. 20. The computer readable storage media of claim 19, wherein the sigmoid function ga(A) is defined by: ga(A)=Ja+Ka1+e-LaA,where Ja, Ka, and La are parameters determined by the training of the LSTM neural network.