Optimizing data center controls using neural networks

1. A method comprising:

receiving a state input characterizing a current state of a facility having one or more computer servers for processing data and mechanical and electrical equipment and controlled by control software;

obtaining a first set of facility setting slates that each define a respective combination of possible facility settings that affect a resource efficiency of the facility;

for each facility setting slate in the first set of facility setting slates:

determining, for each of a plurality of operating constraints for the facility, a prediction of whether the operating constraint would be violated if facility settings defined by the facility setting slate are adopted in response to receiving the state input;

generating a filtered set of facility setting slates that includes only the facility setting slates from the first set of facility setting slates that are predicted to not violate any of the plurality of operating constraints;

generating a plurality of efficiency scores for the filtered set of facility setting slates, the generating comprising: for each facility setting slate in the filtered set of facility setting slates, processing the state input and the facility setting slate through each first deep neural network in an ensemble of first deep neural networks, wherein each first deep neural network in the ensemble comprises one or more hidden layers and an output layer, and wherein each first deep neural network in the ensemble has been configured through training on training data that comprises historical data from past sensor readings from one or more sensors to:

receive the state input and the facility setting slate, and

process the state input and the facility setting slate to generate an efficiency score that characterizes a predicted resource efficiency of the facility if the facility settings defined by the facility setting slate are adopted in response to receiving the state input; and

selecting, based on the plurality of efficiency scores for the facility setting slates in the filtered set of facility setting slates, new values for the facility settings.

2. The method of claim 1 , wherein each first deep neural network in the ensemble has been trained on the training data that comprises historical data from past sensor readings through a reinforcement learning loop that uses a reward signal.

3. The method of claim 1 , wherein for each facility setting slate in the first set of facility setting slates, determining the prediction of whether the operating constraint would be violated comprises:

processing the state input and the facility setting slate to generate a constraint score associated with the operating constraint for the facility, through each second deep neural network of an ensemble of second deep neural networks specific to the operating constraint,

wherein the constraint score characterizes a predicted value of an operating property of the facility if the facility settings defined by the facility setting slate are adopted in response to receiving the state input; and

determining whether the operating constraint would be violated based on determining whether the constraint score associated with the operating constraint satisfies a respective predetermined threshold.

4. The method of claim 3 , wherein generating the constraint score further comprises:

obtaining a respective candidate constraint score from each second deep neural network of the ensemble of second deep neural networks by processing the state input and the facility setting slate using a corresponding second deep neural network; and

generating the constraint score based on a measure of central tendency of the respective candidate constraint scores.

5. The method of claim 1 , wherein selecting, based on the plurality of efficiency scores for the facility setting slates in the filtered set of facility setting slates, new values for the facility settings further comprises:

for each facility setting slate of the filtered set of facility setting slates:

obtaining a respective efficiency score from each first deep neural network of the ensemble of first deep neural networks by processing the state input and the facility setting slate using a corresponding first deep neural network; and

generating a respective optimized efficiency score for the facility setting slate based on a measure of central tendency of the respective efficiency scores; and

selecting, based on the respective optimized efficiency scores, new values for the facility settings.

6. The method of claim 1 , wherein selecting, based on the plurality of efficiency scores for the facility setting slates in the filtered set of facility setting slates, new values for the facility settings further comprises:

for each facility setting slate of the filtered set of facility setting slates, determining a respective optimized efficiency score based on the plurality of efficiency scores obtained for the facility setting slate using the ensemble of first deep neural networks;

selecting a facility setting slate, from the filtered set of facility setting slates, that has a highest optimized efficiency score; and

selecting, as the new values, values represented by the facility settings of the selected facility setting slate.

7. The method of claim 1 , wherein the efficiency score comprises a predicted long-term power usage effectiveness of the facility.

8. The method of claim 2 , wherein each first deep neural network in the ensemble has been re-trained on observed efficiency scores for the facility, wherein the observed efficiency scores have been observed after facility setting slates selected using the ensemble of the first deep neural networks have been adopted.

9. A system comprising:

one or more computers and one or more storage devices on which are stored instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations comprising:

receiving a state input characterizing a current state of a facility having one or more computer servers for processing data and mechanical and electrical equipment and controlled by control software;

obtaining a first set of facility setting slates that each define a respective combination of possible facility settings that affect a resource efficiency of the facility;

for each facility setting slate in the first set of facility setting slates:

determining, for each of a plurality of operating constraints for the facility, a prediction of whether the operating constraint would be violated if facility settings defined by the facility setting slate are adopted in response to receiving the state input;

generating a filtered set of facility setting slates that includes only the facility setting slates from the first set of facility setting slates that are predicted to not violate any of the plurality of operating constraints;

generating a plurality of efficiency scores for the filtered set of facility setting slates, the generating comprising: for each facility setting slate in the filtered set of facility setting slates, processing the state input and the facility setting slate through each first deep neural network in an ensemble of first deep neural networks, wherein each first deep neural network in the ensemble comprises one or more hidden layers and an output layer, and wherein each first deep neural network in the ensemble has been configured through training on training data that comprises historical data from past sensor readings from one or more sensors to:

receive the state input and the facility setting slate, and

process the state input and the facility setting slate to generate an efficiency score that characterizes a predicted resource efficiency of the facility if the facility settings defined by the facility setting slate are adopted in response to receiving the state input; and

selecting, based on the plurality of efficiency scores for the facility setting slates in the filtered set of facility setting slates, new values for the facility settings.

10. The system of claim 9 , wherein each first deep neural network in the ensemble has been trained on the training data that comprises historical data from past sensor readings through a reinforcement learning loop that uses a reward signal.

11. The system of claim 9 , wherein for each facility setting slate in the first set of facility setting slates, determining the prediction of whether the operating constraint would be violated comprises:

processing the state input and the facility setting slate to generate a constraint score associated with the operating constraint for the facility, through each second deep neural network of an ensemble of second deep neural networks specific to the operating constraint,

wherein the constraint score characterizes a predicted value of an operating property of the facility if the facility settings defined by the facility setting slate are adopted in response to receiving the state input; and

determining whether the operating constraint would be violated based on determining whether the constraint score associated with the operating constraint satisfies a respective predetermined threshold.

12. The system of claim 11 , wherein generating the constraint score further comprises:

obtaining a respective candidate constraint score from each second deep neural network of the ensemble of second deep neural networks by processing the state input and the facility setting slate using a corresponding second deep neural network; and

generating the constraint score based on a measure of central tendency of the respective candidate constraint scores.

13. The system of claim 9 , wherein selecting, based on the plurality of efficiency scores for the facility setting slates in the filtered set of facility setting slates, new values for the facility settings further comprises:

for each facility setting slate of the filtered set of facility setting slates:

obtaining a respective efficiency score from each first deep neural network of the ensemble of first deep neural networks by processing the state input and the facility setting slate using a corresponding first deep neural network; and

generating a respective optimized efficiency score for the facility setting slate based on a measure of central tendency of the respective efficiency scores; and

selecting, based on the respective optimized efficiency scores, new values for the facility settings.

14. The system of claim 9 , wherein selecting, based on the plurality of efficiency scores for the facility setting slates in the filtered set of facility setting slates, new values for the facility settings further comprises:

for each facility setting slate of the filtered set of facility setting slates, determining a respective optimized efficiency score based on the plurality of efficiency scores obtained for the facility setting slate using the ensemble of first deep neural networks;

selecting a facility setting slate, from the filtered set of facility setting slates, that has a highest optimized efficiency score; and

selecting, as the new values, values represented by the facility settings of the selected facility setting slate.

15. One or more non-transitory computer readable storage media encoded with instructions that, when executed by one or more computers, cause the one or more computers to perform operations comprising:

receiving a state input characterizing a current state of a facility having one or more computer servers for processing data and mechanical and electrical equipment and controlled by control software;

obtaining a first set of facility setting slates that each define a respective combination of possible facility settings that affect a resource efficiency of the facility;

for each facility setting slate in the first set of facility setting slates:

determining, for each of a plurality of operating constraints for the facility, a prediction of whether the operating constraint would be violated if facility settings defined by the facility setting slate are adopted in response to receiving the state input;

generating a filtered set of facility setting slates that includes only the facility setting slates from the first set of facility setting slates that are predicted to not violate any of the plurality of operating constraints;

generating a plurality of efficiency scores for the filtered set of facility setting slates, the generating comprising: for each facility setting slate in the filtered set of facility setting slates, processing state input and the facility setting slate through each first deep neural network in an ensemble of first deep neural networks, wherein each first deep neural network in the ensemble comprises one or more hidden layers and an output layer, and wherein each first deep neural network in the ensemble has been configured through training on training data that comprises historical data from past sensor readings from one or more sensors to:

receive the state input and the facility setting slate, and

process the state input and the facility setting slate to generate an efficiency score that characterizes a predicted resource efficiency of the facility if the facility settings defined by the facility setting slate are adopted in response to receiving the state input; and

selecting, based on the plurality of efficiency scores for the facility setting slates in the filtered set of facility setting slates, new values for the facility settings.

16. The computer readable storage media of claim 15 , wherein each first deep neural network in the ensemble has been trained on the training data that comprises historical data from past sensor readings through a reinforcement learning loop that uses a reward signal.

17. The computer readable storage media of claim 15 , wherein for each facility setting slate in the first set of facility setting slates, determining the prediction of whether the operating constraint would be violated comprises:

processing the state input and the facility setting slate to generate a constraint score associated with the operating constraint for the facility, through each second deep neural network of an ensemble of second deep neural networks specific to the operating constraint,

wherein the constraint score characterizes a predicted value of an operating property of the facility if the facility settings defined by the facility setting slate are adopted in response to receiving the state input; and

determining whether the operating constraint would be violated based on determining whether the constraint score associated with the operating constraint satisfies a respective predetermined threshold.

18. The computer readable storage media of claim 17 , wherein generating the constraint score further comprises:

obtaining a respective candidate constraint score from each second deep neural network of the ensemble of second deep neural networks by processing the state input and the facility setting slate using a corresponding second deep neural network; and

generating the constraint score based on a measure of central tendency of the respective candidate constraint scores.

19. The computer readable storage media of claim 15 , wherein

selecting, based on the plurality of efficiency scores for the facility setting slates in the filtered set of facility setting slates, new values for the facility settings further comprises:

for each facility setting slate of the filtered set of facility setting slates:

obtaining a respective efficiency score from each first deep neural network of the ensemble of first deep neural networks by processing the state input and the facility setting slate using a corresponding first deep neural network; and

generating a respective optimized efficiency score for the facility setting slate based on a measure of central tendency of the respective efficiency scores; and

selecting, based on the respective optimized efficiency scores, new values for the facility settings.

20. The computer readable storage media of claim 15 , wherein selecting, based on the plurality of efficiency scores for the facility setting slates in the filtered set of facility setting slates, new values for the facility settings further comprises:

for each facility setting slate of the filtered set of facility setting slates, determining a respective optimized efficiency score based on the plurality of efficiency scores obtained for the facility setting slate using the ensemble of first deep neural networks;

selecting a facility setting slate, from the filtered set of facility setting slates, that has a highest optimized efficiency score; and

selecting, as the new values, values represented by the facility settings of the selected facility setting slate.