A Hyperparameter Optimization Algorithm for the LSTM Temperature Prediction Model in Data Center
Algorithm 1
Hyperparameter optimization algorithm based on MLP.
Input: hyperparameter space (X), initial selection of hyperparameter amount (N), screening ratio (t), disturbance ratio (k), stability weight coefficients (W)
Output: a set of optimal hyperparameters
Step:
(1)
x = random_get_hyperparameters (N)//initially randomly select N sets of hyperparameters from the hyperparameter space
(2)
R = run_and_get_loss (A (x))//Train the A model with the selected hyperparameters to get RMSE
(3)
MLP.fit (x, R)//Training MLP model with trained hyperparameters and their RMSE
(4)
While (N > 1) Do
(5)
N = N/t//Gradually reduce N according to the screening ratio
(6)
MLP.predict (x)//Predict the RMSE of hyperparameters with MLP model
(7)
x = use_MLP_get_hyperparameters (N)//
Select the top N sets of hyperparameters according to the MLP prediction results
(8)
x = ++use_disturbance_hyperparameters (N/k)//Perturb the optimal hyperparameters according to a certain proportion to increase the possibility of optimal solutions
(9)
R = run_and_get_loss (A (x))
(10)
MLP.refit (x, R)//retraining MLP
(11)
For i ≤ k DO//Take the first k groups of hyperparameters with the smallest RSME
(12)
L[i] = W∗Normalized_Quadratic loss function [i] +RSME[i]//Normalize the squared loss function
