System and method for model parameter optimization
One embodiment can provide a method and system for tuning parameters of a numerical model of a physical system. During operation, the system can obtain, using a machine-learning technique, a parameter-transform model for mapping parameters of the numerical model at a first resolution to parameters of the numerical model at a second resolution, the second resolution being higher than the first resolution. The system can perform a parameter-tuning operation on the numerical model at a first resolution to obtain a first set of tuned parameters and apply the parameter-transform model on the first set of tuned parameters to obtain a second set of tuned parameters at a second resolution. The system can then generate behavior information associated with the physical system by running the numerical model at the second resolution using the second set of tuned parameters.
1. A computer-implemented method comprising:
obtaining, using a first machine-learning technique, a parameter-transform model for mapping parameters of a numerical model other than the parameter-transform model at a first resolution to parameters of the numerical model at a second resolution, wherein the second resolution is higher than the first resolution, and obtaining the parameter-transform model comprises:
obtaining parameter-transform training samples comprising a set of training parameters of the numerical model associated with the first resolution and a corresponding set of training parameters of the numerical model associated with the second resolution; and
training the parameter-transform model using the obtained parameter-transform training samples;
performing a parameter-tuning operation on the numerical model at the first resolution to obtain a first set of tuned parameters;
applying the parameter-transform model on the first set of tuned parameters to obtain a second set of tuned parameters at the second resolution;
generating behavior information associated with the physical system by running the numerical model at the second resolution using the second set of tuned parameters; and
obtaining, using a second machine-learning technique, a parameter-mapping model for mapping an output of the numerical model to actual parameters used for running the numerical model at the first resolution.
2. The computer-implemented method of claim 1 , wherein the set of training parameters of the numerical model associated with the first resolution is obtained by applying the parameter-mapping model to outputs of the numerical model running at the second resolution; and wherein the set of training parameters associated with the second resolution comprises actual parameters used for generating the outputs by executing the numerical model at the second resolution.
3. The computer-implemented method of claim 1 , wherein the parameter-mapping model comprises a deep neural network.
4. The computer-implemented method of claim 1 , further comprising:
obtaining parameter-mapping training samples, which comprise parameters within a predetermined parameter space and outputs of the numerical model generated by executing the numerical model at the first resolution using the parameters within the predetermined parameter space; and
training the parameter-mapping model using the obtained parameter-mapping training samples.
5. The computer-implemented method of claim 4 , wherein obtaining the parameter-mapping training samples further comprises up-mapping the outputs of the numerical model from the first resolution to the second resolution.
6. The computer-implemented method of claim 4 , wherein obtaining the parameter-mapping training samples further comprises preprocessing outputs of the numerical model to facilitate the training of the parameter-mapping model.
7. A computer system comprising:
a processor; and
a storage device coupled to the processor and storing instructions which when executed by the processor cause the processor to perform a method, the method comprising:
obtaining, using a first machine-learning technique, a parameter-transform model for mapping parameters of a numerical model other than the parameter-transform model at a first resolution to parameters of the numerical model at a second resolution, wherein the second resolution is higher than the first resolution, and obtaining the parameter-transform model comprises:
obtaining parameter-transform training samples comprising a set of training parameters of the numerical model associated with the first resolution and a corresponding set of training parameters of the numerical model associated with the second resolution; and
training the parameter-transform model using the obtained parameter-transform training samples;
performing a parameter-tuning operation on the numerical model at the first resolution to obtain a first set of tuned parameters; and
applying the parameter-transform model on the first set of tuned parameters to obtain a second set of tuned parameters at the second resolution;
generating behavior information associated with the physical system by running the numerical model at the second resolution using the second set of tuned parameters; and
obtaining, using a second machine-learning technique, a parameter-mapping model for mapping an output of the numerical model to actual parameters used for running the numerical model at the first resolution.
8. The computer system of claim 7 , wherein the set of training parameters of the numerical model associated with the first resolution is obtained by applying the parameter-mapping model to outputs of the numerical model running at the second resolution; and wherein the set of training parameters associated with the second resolution comprises actual parameters used for generating the outputs by executing the numerical model at the second resolution.
9. The computer system of claim 7 , wherein the parameter-mapping model comprises a deep neural network.
10. The computer system of claim 7 , wherein the method further comprises:
obtaining parameter-mapping training samples, which comprise parameters within a predetermined parameter space and outputs of the numerical model generated by executing the numerical model at the first resolution using the parameters within the predetermined parameter space; and
training the parameter-mapping model using the obtained parameter-mapping training samples.
11. The computer system of claim 10 , wherein obtaining the parameter-mapping training samples further comprises up-mapping the outputs of the numerical model from the first resolution to the second resolution.
12. The computer system of claim 10 , wherein obtaining the parameter-mapping training samples further comprises preprocessing outputs of the numerical model to facilitate the training of the parameter-mapping model.
13. A computer-implemented method comprising:
obtaining, using a machine-learning technique, a parameter-transform model for mapping parameters of a numerical model other than the parameter-transform model at a first resolution to parameters of the numerical model at a second resolution, wherein the second resolution is higher than the first resolution, the numerical model comprises a deep regression network, and the parameter-transform model comprises a gradient-boosting regression model;
performing a parameter-tuning operation on the numerical model at the first resolution to obtain a first set of tuned parameters;
applying the parameter-transform model on the first set of tuned parameters to obtain a second set of tuned parameters at the second resolution; and
generating behavior information associated with the physical system by running the numerical model at the second resolution using the second set of tuned parameters.