Generation of a control system for a target system
The invention relates to a method for generating a control system for a target system, wherein: operational data is received; a first neural model component is trained with the received operational data for generating a prediction on a state of the target system based on the received operational data; a second neural model component is trained with the operational data for generating a regularizer for use in inverting the first neural model component; and the control system is generated by inverting the first neural model component by optimization and arranging to apply the regularizer generated with the second neural model component in the optimization. The invention relates also to a system and a computer program product.
1. A method for generating a control system for a target system, the method comprising:
receiving operational data of at least one source system,
training a first neural model component of a neural network with the received operational data, wherein the first neural model component is trained to generate a prediction on a state of the target system based on the received operational data,
training a second neural model component of the neural network with the operational data, wherein the second neural model component is trained to generate a regularization term, and
generating the control system for the target system by inverting the first neural model component by optimization, wherein the regularization term generated with the second neural model component is applied in the optimization to direct the optimization toward values that correspond to typical samples in the operational data,
wherein generating the control system for the target system comprises using a value function in the optimization and adding the regularization term to the value function as a reward.
2. The method of claim 1 , wherein the second neural model component is one of the following: a denoising neural network, a generative adversarial network, a variational autoencoder, a ladder network.
3. The method of claim 1 , wherein the control system is generated by optimization by generating a policy model component configured to generate at least one control signal and optimizing parameters of the policy model component.
4. The method of claim 3 , wherein the policy model is implemented with one of the following: with a trained third neural model component, by optimizing parameters of a controller component belonging to the control system.
5. A control system for a target system, the control system comprising at least one neural network, wherein the control system is configured to be generated by:
receiving operational data of at least one source system,
training a first neural model component of the neural network with the received operational data, wherein the first neural model component is trained to generate a prediction on a state of the target system based on the received operational data in response to an input signal,
training a second neural model component of the neural network with the operational data, wherein the second neural model component is trained to generate a regularization term, and
generating the control system for the target system by inverting the first neural model component by optimization, wherein the regularization term generated with the second neural model component is applied in the optimization to direct the optimization toward values that correspond to typical samples in the operational data,
wherein generating the control system for the target system comprises using a value function in the optimization and adding the regularization term to the value function as a reward.
6. The system of claim 5 , wherein the second neural model component is one of the following: a denoising neural network, a generative adversarial network, a variational autoencoder, a ladder network.
7. The system of claim 5 , wherein the control system is configured to be generated by optimization by generating a policy model component configured to generate at least one control signal and optimizing parameters of the policy model component.
8. The system of claim 7 , wherein the policy model is implemented with one of the following: with a trained third neural model component, by optimizing parameters of a controller component belonging to the control system.
9. A non-transitory computer program product comprising at least one computer-readable media having computer-executable program code instructions stored therein, for generating a control system for a target system, wherein the computer-executable program code instructions when executed on a computer cause the computer to:
receive operational data of at least one source system,
train a first neural model component of a neural network with the received operational data, wherein the first neural model component is trained to generate a prediction on a state of the target system based on the received operational data,
train a second neural model component of the neural network with the operational data, wherein the second neural model component is trained to generate a regularization term, and
generate the control system for the target system by inverting the first neural model component by optimization, wherein the regularization term generated with the second neural model component is applied in the optimization to direct the optimization toward values that correspond to typical samples in the operational data,
wherein the instructions to generate the control system comprises instructions to use a value function in the optimization and add the regularization term to the value function as a reward.
10. The computer program product of claim 9 , wherein the second neural model component is one of the following: a denoising neural network, a generative adversarial network, a variational autoencoder, a ladder network.
11. The computer program product of claim 9 , wherein the control system is generated by optimization by generating a policy model component configured to generate at least one control signal and optimizing parameters of the policy model component.
12. The computer program product of claim 11 , wherein the policy model is implemented with one of the following: with a trained third neural model component, by optimizing parameters of a controller component belonging to the control system.
13. The method of claim 1 , wherein inverting the first neural model component refers to a procedure in which inputs or control actions to the target system that result in desired outputs or states of the target system are determined.
14. The system of claim 5 , wherein inverting the first neural model component refers to a procedure in which inputs or control actions to the target system that result in desired outputs or states of the target system are determined.
15. The computer program product of claim 9 , wherein inverting the first neural model component refers to a procedure in which inputs or control actions to the target system that result in desired outputs or states of the target system are determined.
16. The method of claim 1 , wherein the second neural model component is trained to generate a regularization term by training the second neural model component to denoise samples of the operational data.
17. The method of claim 16 , wherein the second neural model component is a denoising autoencoder.
18. The system of claim 5 , wherein the second neural model component is trained to generate a regularization term by training the second neural model component to denoise samples of the operational data.
19. The system of claim 18 , wherein the second neural model component is a denoising autoencoder.
20. The computer program product of claim 9 , wherein the second neural model component is trained to generate a regularization term by training the second neural model component to denoise samples of the operational data.
21. The computer program product of claim 20 , wherein the second neural model component is a denoising autoencoder.