Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Continued Examination Under 37 CFR 1.114
A request for continued examination under 37 CFR 1.114, including the fee set forth in 37 CFR 1.17(e), was filed in this application after final rejection. Since this application is eligible for continued examination under 37 CFR 1.114, and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114. Applicant's submission filed on 09/10/2026 has been entered.
Status of Claims
This action is in reply to the amendments and remarks filed on 09/10/2026.
Claims 1-20 are pending.
Claims 1, 3-4, 9, 13, 15-16, and 20 have been amended.
Response to Arguments
Applicant’s arguments, with respect to the rejection(s) of claim(s) 1 and 13 under 35 U.S.C. 103, have been considered but they are not persuasive. Applicant argues that no reference teaches the amended limitations now stating “for an identified change to the relationship between the energy usage and the associated characteristics ... collecting, based on the identified change, third data for a third time period; and generating a second machine-trained model based on the third data collected after the identified change, wherein the second machine-trained model replaces the first machine-trained model so as to avoid using data associated with a previous relationship between the energy usage and the associated characteristics”, since Li's “offline parameter-selection process is fundamentally different from identifying a change in a deployed relationship that triggers collecting new data and generating a replacement model”. The examiner respectfully disagrees.
Due to the broadness of the claim language, it is maintained that Li teaches the argued limitations, since section 5D teaches consecutively computing different designs of the neural network including a trained LSTM output states and actions (third data collected after the identified change) are input into an iteratively trained Q network for further turning as a new “training episode” (generating a second machine-trained model). Further, sections 5B-D teach iteratively updating the trained Q network for further tuning as a new “training episode” (second machine-trained model replaces the first machine-trained model) to “replace the original neural network”. Here, using the trained model on state-action pairs of temperature and load factor data over time to compute validation errors for striking “a balance between minimizing PUE and preventing overheating in the server zone” that the equipment is operating; thus, identifying a relationship between workload energy use and resulting ambient temperature as well as the change between each, and maintained as teaching the claimed scope.
See 35 U.S.C 103 section for full mapping of claim limitations necessitated by applicant amendments.
Applicant’s arguments, with respect to the rejection(s) of claim(s) 1 and 13 under 35 U.S.C. 103, have been considered but they are not persuasive. Applicant argues that no reference teaches the amended limitations now stating “wherein the first machine-trained model accounts for a dependency between the energy usage and the outside air temperature to distinguish changes to the energy usage attributable to the outside air temperature from changes not explainable by the outside air temperature”, since Li's “offline parameter-selection process is fundamentally different from identifying a change in a deployed relationship that triggers collecting new data and generating a replacement model” and Stein does not remedy this issue. The examiner respectfully disagrees.
The combination has been found to teach the amended limitations, since Li teaches training the model on state-action pairs of temperature and load factor data for striking "a balance between minimizing PUE and preventing overheating in the server zone" that the equipment is operating; thus, identifying the relationship between workload energy use and resulting ambient temperature as well as using “load and weather information” for datacenter zone efficiency predictions. Further, Stein is found to teach in combination paragraphs 0113-0114, 0120-0122, and 0152-0153 teach giving recommendations for curtailment to “end users” for resolving “issues” between factors including energy usage and outside weather data for a “data center”.
Further still, Stein cited paragraphs teach processing energy use and weather dependent data utilizing probability distributions for a data center, wherein paragraphs 0093-0097 further define the probability distribution used based on forecast error computations; thus, reading on the claimed scope.
See 35 U.S.C 103 section for full mapping of claim limitations necessitated by applicant amendments.
Applicant’s arguments, with respect to the rejection(s) of claim(s) 3 and 15 under 35 U.S.C. 103, have been considered but they are not persuasive. Applicant argues that no reference teaches the amended limitations now stating “training the second machine-trained model using differenced data comprising a difference between a data feature vector associated with a time step and a data feature vector associated with a previous time step”, since Baig’s “adaptive sliding windows operate on raw resource-utilization data and do not teach or suggest training on such differenced feature vectors”. The examiner respectfully disagrees.
Due to the broadness of the claim language, Baig has been found to teach the limitations, since sections 6.3 and 7 teach different window sizes covering different time periods (difference between a data feature vector associated with a time step and a data feature vector associated with a previous time step) for collecting a dataset for iteratively training the model (generating the second machine-trained model based on the third data) and iteratively determining error minimization.
See 35 U.S.C 103 section for full mapping of claim limitations necessitated by applicant amendments.
Applicant’s arguments, with respect to the rejection(s) of claim(s) 4 and 16 under 35 U.S.C. 103, have been considered but they are not persuasive. Applicant argues that no reference teaches the amended limitations now stating “wherein the control limits comprise an upper control limit calculated as the EWMA plus the EWMSTD times a scaling factor and a lower control limit calculated as the EWMA minus the EWMSTD times the scaling factor, and wherein the scaling factor is configured by an administrator based on a desired sensitivity of a change-detection operation for detecting changes to the relationship between the energy usage and the associated characteristics”, since “Cambron's EWMA power-curve control charts do not teach a configurable scaling factor tied to change-detection sensitivity in this dependency-aware modeling context”. The examiner respectfully disagrees.
Due to the broadness of the claim language, it is maintained that the combination teaches the limitations, since Cambron section 3 and Equations 4-5 teach upper and lower limit calculations including adding/subtracting a control limit, wherein “We choose k = 3, corresponding to a 3σ control limits” (EWMSTD times a scaling factor…configured by an administrator)”.
See 35 U.S.C 103 section for full mapping of claim limitations necessitated by applicant amendments.
Applicant’s arguments, with respect to the rejection(s) of claim(s) 9-12 and 20 under 35 U.S.C. 103, have been considered but they are not persuasive. Applicant argues that no reference teaches the amended limitations now stating “wherein the identified change is further based on the difference being one of greater than the first value or less than the second value at least a threshold number of times, and wherein the threshold number of times is configured based on a desired sensitivity of a change-detection operation for detecting changes to the relationship between the energy usage and the associated characteristics”, since “Neither Biag nor Cambron” teach a “violation-count requirement in combination with the dependency-aware model”. The examiner respectfully disagrees.
Due to the broadness of the claim language, Cambron has been found to teach the limitations, since sections 3, 5.1, and Fig. 1 teach the EWMA target value (identified change) is greater than and less than the calculated EWMA bounds across different zones over observation times (based on the difference being one of greater than the first value or less than the second value at least a threshold number of times), including “80 days” of detection data and “number of runs required to reach convergence of the ARL” (threshold number of times is configured based on a desired sensitivity of a change-detection operation).
See 35 U.S.C 103 section for full mapping of claim limitations necessitated by applicant amendments.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The factual inquiries set forth in Graham v. John Deere Co., 383 U.S. 1, 148 USPQ 459 (1966), that are applied for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention.
Claims 1-8 and 13-19 are rejected under 35 U.S.C. 103 as being unpatentable over Li et al (“Transforming Cooling Optimization for Green Data Center via Deep Reinforcement Learning”, 2018) hereinafter Li, in view of Stein et al (US Pub 20140058572) hereinafter Stein.
Regarding claims 1 and 13, Li teaches a method comprising; and an apparatus comprising: a memory: and at least one processor coupled to the memory and, based at least in part on information stored in the memory (sections 5-6 teach using a “CPU” for executing the embodiments of the disclosure, known to be included in a computer system and communicatively coupled to one or memories), the at least one processor is configured to:
receiving, from a set of sensors associated with a datacenter, a data stream comprising measurements of energy usage for, and associated characteristics of, the datacenter, wherein the associated characteristics comprise at least an outside air temperature, (sections 3, 6, and Table 3 teach “We are given a time-varying tuple of the ambient air temperature Tamb and the load factor Hite” (associated characteristics comprise at least an outside air temperature), and multiple PCU temperatures (associated characteristics), power consumptions (energy usage), and air/water flow rate readings (first and second energy usage data). “With the above data, we utilize the proposed algorithm to train the Q and μ network. In this case, as the power consumption can be directly computed by the fan law from the airflow rate, we will only rely on the Q to approximate the inlet temperature which we use as the thermal indicator (modeling a relationship).”);
collecting, from the data stream, first data for a first time period and second data for a second time period wherein the first data and the second data are associated with rolling windows of the data stream (sections 5-5A, 6, and Table 3 teach PCU temperatures, power consumptions, and air/water flow rate readings; and further “We collected these data entries for every 3 minutes (rolling windows of the data stream) from March 1 to 15 of 2017. For these data, we use the first 85% as the training data (first data for a first time period) and the last 15% as the test data (second data for a second time period)”);
generating, based on the first data for the first time period, a first machine-trained model modeling a relationship between the energy usage and the associated characteristics (section 6 teaches “With the above data (based on the first data for the first time period), we utilize the proposed algorithm to train the Q and μ network. In this case, as the power consumption can be directly computed by the fan law from the airflow rate, we will only rely on the Q (first machine-trained model) to approximate the inlet temperature which we use as the thermal indicator (modeling a relationship).”), wherein the first machine-trained model accounts for a dependency between the energy usage and the outside air temperature to distinguish changes to the energy usage attributable to the outside air temperature from changes not explainable by the outside air temperature (sections 3-4 and section 6, paragraphs 1-4 teach training the Q network (first machine-trained model) on state data including the ambient temperature and load factor variables (between the energy usage and the outside air temperature) to compute a cost of state-action pair in order to minimize the “optimization problem” (accounts for a dependency)); and
for an identified change to the relationship between the energy usage and the associated characteristics based on a first prediction error associated with the second time period that measures a difference between a first predicted energy usage for the second time period based on the first machine-trained model and a first actual energy usage for the second time period indicated in the second data being one of greater than a first value or less than a second value (sections 4D, 6, and Algorithm 1 teach using a validation set for computing “validation error” (first prediction error associated with the second time period) of the model; wherein “Q outputs the predicted energy and temperature data, concatenated as yr”, and “minimizing the error between the predicted yr (measures a difference between a first predicted energy usage for the second time period based on the first machine-trained model) and the real data (and a first actual energy usage for the second time period indicated in the second data)” used in the error computation that is further compared to predetermined values, “EQval” (being one of greater than a first value or less than a second value)):
collecting, based on the identified change, third data for a third time period (section 5D teaches “different designs of the neural network” including a trained LSTM output states and actions (third data) are input into a trained Q network); and
generating a second machine-trained model based on the third data collected after the identified change (section 5D teaches consecutively computing different designs of the neural network including a trained LSTM output states and actions (third data collected after the identified change) are input into an iteratively trained Q network for further turning as a new “training episode” (generating a second machine-trained model)),
wherein the second machine-trained model replaces the first machine-trained model so as to avoid using data associated with a previous relationship between the energy usage and the associated characteristics (section 5B-D teach iteratively updating the trained Q network for further tuning as a new “training episode” (second machine-trained model replaces the first machine-trained model) in order to “replace the original neural network”).
However, while Li teaches using “load and weather information” for datacenter zone efficiency predictions, Li does not teach wherein the outside air temperature is a temperature of air external to the datacenter, and displaying, to a datacenter operator via an output interface, an indication of the identified change and a recommendation for remediation of the identified change.
Stein teaches wherein the outside air temperature is a temperature of air external to the datacenter, and displaying, to a datacenter operator via an output interface, an indication of the identified change and a recommendation for remediation of the identified change (paragraphs 0113-0114, 0120-0122, and 0152-0153 teach giving recommendations for curtailment to “end users” for resolving “issues” between factors including energy usage and outside weather data for a “data center”).
Thus, it would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to implement Stein’s teachings of datacenter resource utilization results and recommendations for fixing poor results output to a user operator into Li‘s teaching of using machine learning algorithms for optimizing datacenter power usage and error metric calculations in order to maximize “savings and make appropriate recommendations to optimize the return on the building manager's effort” (Stein, paragraph 0120).
Regarding claims 2 and 14, the combination of Li and Stein teach all the claim limitations of claims 1 and 13 above; and further teach wherein the generating the second machine-trained model is further based on the second data; and
using the second machine-trained model to predict energy usage for a subsequent time period following the third time period to continue monitoring the datacenter for subsequent changes to the relationship between the energy usage and the associated characteristics (Li, section 5D teaches “different designs of the neural network” including a trained LSTM output states and actions (third data) are iteratively input into a trained Q network for further tuning as a new “training episode” (generating the second machine-trained model); wherein sections 4D, 6, and Algorithm 1 teach the Q network primarily trained and validated with “predicted yr and the real data (second data)” for predicting “energy and temperature data, concatenated as yr”, and “minimizing the error between the predicted yr and the real data”).
Regarding claims 5 and 17, the combination of Li and Stein teach all the claim limitations of claims 1 and 13 above; and further teach wherein the associated characteristics comprise at least the outside air temperature and the energy usage comprises at least a first energy usage data associated with a first power consumed by equipment providing information technology (IT) functions at the datacenter and a second energy usage data associated with a second power consumed by the datacenter, wherein the relationship between the energy usage and the associated characteristics comprises a particular relationship between the second power, the first power, and the associated characteristics (Li, sections 3, 6, and Table 3 teach “We are given a time-varying tuple of the ambient air temperature Tamb and the load factor Hite” (outside air temperature), and multiple power consumptions (first and second energy usage data). “With the above data, we utilize the proposed algorithm to train the Q and μ network. In this case, as the power consumption can be directly computed by the fan law from the airflow rate, we will only rely on the Q to approximate the inlet temperature which we use as the thermal indicator (modeling a relationship).”).
Regarding claim 6, the combination of Li and Stein teach all the claim limitations of claim 5 above; and further teach wherein the particular relationship between the second power, the first power, and the associated characteristics comprises a function for calculating a power usage effectiveness (PUE) based on the first power and the associated characteristics, wherein the PUE is calculated by dividing the second power by the first power (Li, sections 4-5 teach error calculations involving the different powers and temperatures, and further calculating PUE from total power divided by equipment power).
Regarding claims 7 and 18, the combination of Li and Stein teach all the claim limitations of claims 5 and 17 above; and further teach wherein the first energy usage data and the second energy usage data comprise one or more of energy usage data at a first set of two or more levels of granularity in space or energy usage data at a second set of two or more levels of granularity in time, wherein the first set of two or more levels of granularity in space comprises one or more of an IT device-level granularity, a rack-level granularity, a group-of-racks level granularity, a room level granularity, a group-of-rooms level granularity, a floor level granularity, a building level granularity, or a datacenter level granularity, wherein the second set of two or more levels of granularity in time comprises one or more of seconds, minutes, hours, days, weeks, months, quarters, or years (Li, sections 5-5A, 6, and Table 3 teach PCU temperatures, power consumptions per rack grouping (first energy usage data and the second energy usage data comprise one or more of energy usage data at a first set of two or more levels of granularity in space/group-of-racks level granularity), and air/water flow rate readings; and further “We collected these data entries for every 3 minutes (second set of two or more levels of granularity in time comprises one or more of…minutes) from March 1 to 15 of 2017. For these data, we use the first 85% as the training data and the last 15% as the test data”).
Regarding claims 8 and 19, the combination of Li and Stein teach all the claim limitations of claims 7 and 18 above; and further teach receiving a selection of a first level of granularity in time and a second level of granularity in space, wherein generating the first machine-trained model is further based on the first level of granularity in time and the second level of granularity in space (Li, sections 5-5A, 6, and Table 3 teach PCU temperatures, power consumptions per rack grouping (second level of granularity in space), and air/water flow rate readings; and further “We collected these data entries for every 3 minutes (selection of a first level of granularity in time) from March 1 to 15 of 2017. For these data, we use the first 85% as the training data and the last 15% as the test data” to train the Q network).
Claims 3 and 15 are rejected under 35 U.S.C. 103 as being unpatentable over Li et al (“Transforming Cooling Optimization for Green Data Center via Deep Reinforcement Learning”, 2018) hereinafter Li, in view of Stein et al (US Pub 20140058572) hereinafter Stein, in view of Baig et al (“Adaptive sliding windows for improved estimation of data center resource utilization”, 2019) hereinafter Baig.
Regarding claims 3 and 15, the combination of Li and Stein teach all the claim limitations of claims 1 and 13 above; however, the combination does not explicitly teach refraining from using the first machine-trained model until the second machine-trained model is generated based on the third data, wherein the third time period comprises at least a threshold amount of time for collecting data to generate the second machine-trained model after the identified change, and wherein generating the second machine-trained model based on the third data comprises training the second machine-trained model using differenced data comprising a difference between a data feature vector associated with a time step and a data feature vector associated with a previous time step.
Baig teaches refraining from using the first machine-trained model until the second machine-trained model is generated based on the third data (Baig, sections 4 and 6.3 teach using different observation window sizes to train machine learning models (until the second machine-trained model is generated based on the third data) without using the MLP to select the window sizes for data collection in Experiments 1-3 (refraining from using the first machine-trained model) until Experiment 4), wherein the third time period comprises at least a threshold amount of time for collecting data to generate the second machine-trained model after the identified change (Baig, sections 6.3 and 7 teach different window sizes covering different time periods (third time period comprises at least a threshold amount of time for collecting data) for collecting a dataset for iteratively training the model (to generate the second machine-trained model) and iteratively determining error minimization (after the identified change)), and wherein generating the second machine-trained model based on the third data comprises training the second machine-trained model using differenced data comprising a difference between a data feature vector associated with a time step and a data feature vector associated with a previous time step (Baig, sections 6.3 and 7 teach different window sizes covering different time periods (difference between a data feature vector associated with a time step and a data feature vector associated with a previous time step) for collecting a dataset for iteratively training the model (generating the second machine-trained model based on the third data) and iteratively determining error minimization).
Thus, it would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to modify Li‘s teaching of using machine learning algorithms for optimizing datacenter power usage and error metric calculations, as modified by Stein’s teachings of datacenter resource utilization results and recommendations for fixing poor results output to a user operator, to include prediction result observation plots in user VMs for datacenter resource utilization as taught by Baig in order to “improve prediction accuracy” (Baig, section 8).
Claims 4 and 16 are rejected under 35 U.S.C. 103 as being unpatentable over Li et al (“Transforming Cooling Optimization for Green Data Center via Deep Reinforcement Learning”, 2018) hereinafter Li, in view of Stein et al (US Pub 20140058572) hereinafter Stein, in view of Cambron et al (“Power curve monitoring using weighted moving average control charts”, 2019) hereinafter Cambron.
Regarding claims 4 and 16, the combination of Li and Stein teach all the claim limitations of claims 1 and 13 above; and further teach wherein the identified change is further based on the difference being one of greater than the first value or less than the second value at least a threshold number of times (Li, sections 4D, 6, and Algorithm 1 teach using a validation set for computing “validation error” of the model; wherein “Q outputs the predicted energy and temperature data, concatenated as yr”, and “minimizing the error between the predicted yr and the real data (difference)” used in the error computation that is further compared to a predetermined value, “EQval” at least once (being one of greater than the first value or less than the second value at least a threshold number of times)), and wherein the second machine-trained model is generated to replace the first machine-trained model after the identified change (Li, sections 5B and D teach “we train the same evaluation network like CCA to replace the original neural network designed for modeling chiller efficiency”; wherein “different designs of the neural network” including a trained LSTM output states and actions are iteratively input into a trained Q network for further tuning as a new “training episode” (second machine-trained model is generated to replace the first machine-trained model after the identified change). Further, sections 4D, 6, and Algorithm 1 teach the Q network primarily trained and validated with “predicted yr and the real data (second data)”.), wherein the control limits comprise an upper control limit calculated as the EWMA plus the EWMSTD times a scaling factor and a lower control limit calculated as the EWMA minus the EWMSTD times the scaling factor, and wherein the scaling factor is configured by an administrator based on a desired sensitivity of a change-detection operation for detecting changes to the relationship between the energy usage and the associated characteristics.
However, the combination does not explicitly teach and wherein the first value and the second value are control limits derived from an exponentially weighted moving average (EWMA) and an exponentially weighted moving standard deviation (EWMSTD) of prediction errors from a validation time period preceding the second time period, and wherein the control limits comprise an upper control limit calculated as the EWMA plus the EWMSTD times a scaling factor and a lower control limit calculated as the EWMA minus the EWMSTD times the scaling factor, and wherein the scaling factor is configured by an administrator based on a desired sensitivity of a change-detection operation for detecting changes to the relationship between the energy usage and the associated characteristics.
Cambron teaches and wherein the first value and the second value are control limits derived from an exponentially weighted moving average (EWMA) and an exponentially weighted moving standard deviation (EWMSTD) of prediction errors from a validation time period preceding the second time period (sections 3 and 4.3 teach “EWMA” upper and lower control limit exponential calculations including “We choose k = 3, corresponding to a 3σ control limits” (EWM standard deviation); wherein the upper limit is an addition calculation and the lower limit is a subtraction calculation; further the data for validating the model being a time period before the “EWMA calculation”), and
wherein the control limits comprise an upper control limit calculated as the EWMA plus the EWMSTD times a scaling factor and a lower control limit calculated as the EWMA minus the EWMSTD times the scaling factor, and wherein the scaling factor is configured by an administrator based on a desired sensitivity of a change-detection operation for detecting changes to the relationship between the energy usage and the associated characteristics (section 3 and Equations 4-5 teach upper and lower limit calculations including adding/subtracting a control limit, wherein “We choose k = 3, corresponding to a 3σ control limits” (EWMSTD times a scaling factor…configured by an administrator)).
Thus, it would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to modify Li‘s teaching of using machine learning algorithms for optimizing datacenter power usage and error metric calculations, as modified by Stein’s teachings of datacenter resource utilization results and recommendations for fixing poor results output to a user operator, to include EMWA and standard deviation calculations for prediction error analysis as taught by Cambron in order to reduce model “underperformances” and accomplish more accurate prediction accuracy (Cambron, sections 3-4.3).
Claims 9-12 and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Li et al (“Transforming Cooling Optimization for Green Data Center via Deep Reinforcement Learning”, 2018) hereinafter Li, in view of Stein et al (US Pub 20140058572) hereinafter Stein, in view of Baig et al (“Adaptive sliding windows for improved estimation of data center resource utilization”, 2019) hereinafter Baig, in view of Cambron et al (“Power curve monitoring using weighted moving average control charts”, 2019) hereinafter Cambron.
Regarding claim 9, the combination of Li and Stein teach all the claim limitations of claim 1 above; however, the combination does not explicitly teach collecting fourth data for a fourth time period following the first time period and preceding the second time period; determining an average of a second prediction error for the fourth time period based on a second predicted energy usage for the fourth time period predicted by the first machine-trained model and a second actual energy usage for the fourth time period indicated in the fourth data; and determining a standard deviation of the second prediction error, wherein the first value and the second value are based on the average of the second prediction error and the standard deviation of the second prediction error, wherein the identified change is further based on the difference being one of greater than the first value or less than the second value at least a threshold number of times, and wherein the threshold number of times is configured based on a desired sensitivity of a change-detection operation for detecting changes to the relationship between the energy usage and the associated characteristics.
Baig teaches collecting fourth data for a fourth time period following the first time period and preceding the second time period (sections 3-4 and 6.3 teach a sliding window for collecting training time period datasets in a time-series that are each input into the model; and “For validation purposes, the final data-set is randomly split 80/20 for training vs. testing subsets (preceding the second time period)”);
determining an average of a second prediction error for the fourth time period based on a second predicted energy usage for the fourth time period predicted by the first machine-trained model and a second actual energy usage for the fourth time period indicated in the fourth data (section 6.1.2-6.2 and 7 teach determining “MSE” for the different window training sets (fourth time period) for the “true and estimated values” of “resource estimation prediction” via the prediction model);
Thus, it would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to modify Li‘s teaching of using machine learning algorithms for optimizing datacenter power usage and error metric calculations, as modified by Stein’s teachings of datacenter resource utilization results and recommendations for fixing poor results output to a user operator, to include prediction result observation plots in user VMs for datacenter resource utilization as taught by Baig in order to “improve prediction accuracy” (Baig, section 8).
However, the Baig does not explicitly teach determining a standard deviation of the second prediction error, wherein the first value and the second value are based on the average of the second prediction error and the standard deviation of the second prediction error, wherein the identified change is further based on the difference being one of greater than the first value or less than the second value at least a threshold number of times, and wherein the threshold number of times is configured based on a desired sensitivity of a change-detection operation for detecting changes to the relationship between the energy usage and the associated characteristics.
Cambron teaches determining a standard deviation of the second prediction error, wherein the first value and the second value are based on the average of the second prediction error and the standard deviation of the second prediction error (sections 3-4.2 and 5.1 teach “EWMA” of the predictions, computing the standard deviation of the error, and EWMA upper and lower control limit (the first value and the second value) using the EWMA and standard deviation), wherein the identified change is further based on the difference being one of greater than the first value or less than the second value at least a threshold number of times, and wherein the threshold number of times is configured based on a desired sensitivity of a change-detection operation for detecting changes to the relationship between the energy usage and the associated characteristics (sections 3, 5.1, and Fig. 1 teach the EWMA target value (identified change) is greater than and less than the calculated EWMA bounds across different zones over observation times (based on the difference being one of greater than the first value or less than the second value at least a threshold number of times), including “80 days” of detection data and “number of runs required to reach convergence of the ARL” (threshold number of times is configured based on a desired sensitivity of a change-detection operation)).
Thus, it would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to modify Li‘s teaching of using machine learning algorithms for optimizing datacenter power usage and error metric calculations, as modified by Stein’s teachings of datacenter resource utilization results and recommendations for fixing poor results output to a user operator, as modified by Baig’s teachings of prediction result observation plots in user VMs for datacenter resource utilization, to include EMWA and standard deviation calculations for prediction error analysis as taught by Cambron in order to reduce model “underperformances” and accomplish more accurate prediction accuracy (Cambron, sections 3-4.2).
Regarding claim 10, the combination of Li and Stein teach all the claim limitations of claim 9 above; however, the combination does not explicitly teach wherein the average of the second prediction error is an exponentially weighted moving average (EWMA) and the standard deviation of the second prediction error is an exponentially weighted moving standard deviation (EWM standard deviation), wherein the first value is an upper control limit (UCL) calculated as the EWMA plus the EWM standard deviation times a scaling factor and the second value is a lower control limit (LCL) calculated as the EWMA minus the EWM standard deviation times the scaling factor, wherein the scaling factor is configured based on a desired sensitivity of a change-detection operation for detecting changes to the relationship between the energy usage and the associated characteristics of the datacenter.
Cambron teaches wherein the average of the second prediction error is an exponentially weighted moving average (EWMA) and the standard deviation of the second prediction error is an exponentially weighted moving standard deviation (EWM standard deviation), wherein the first value is an upper control limit (UCL) calculated as the EWMA plus the EWM standard deviation times a scaling factor and the second value is a lower control limit (LCL) calculated as the EWMA minus the EWM standard deviation times the scaling factor, wherein the scaling factor is configured based on a desired sensitivity of a change-detection operation for detecting changes to the relationship between the energy usage and the associated characteristics of the datacenter (section 3-section 4, paragraph 1 teach “EWMA” upper and lower control limit exponential calculations including “We choose k = 3 (scaling factor), corresponding to a 3σ control limits” (EWM standard deviation); wherein the upper limit is an addition calculation and the lower limit is a subtraction calculation. Further, “The main principle of the proposed method is to compare measured power to expected power corresponding to measured wind speed with the use of a model. Residual power is then monitored using an EWMA or GWMA chart”.).
Thus, it would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to modify Li‘s teaching of using machine learning algorithms for optimizing datacenter power usage and error metric calculations, as modified by Stein’s teachings of datacenter resource utilization results and recommendations for fixing poor results output to a user operator, as modified by Baig’s teachings of prediction result observation plots in user VMs for datacenter resource utilization, to include EMWA and standard deviation calculations for prediction error analysis as taught by Cambron in order to reduce model “underperformances” and accomplish more accurate prediction accuracy (Cambron, sections 3-4.2).
Regarding claim 11, the combination of Li and Stein teach all the claim limitations of claim 9 above; and further teach collecting fifth data for a fifth time period following the first time period and preceding the fourth time period (Baig, section 6.1.2-6.2 and 7 teach determining “MSE” for the different window training sets (fifth time period) for the “true and estimated values” of “resource estimation prediction” via the prediction model);
determining at least an additional average or an additional standard deviation for a third prediction error based on a third predicted energy usage for the fifth time period predicted by the first machine-trained model and a third actual energy usage for the fifth time period indicated in the fifth data (Baig, section 6.1.2-6.2 and 7 teach determining “MSE” (additional average) for the different window training sets (fifth time period) for the “true and estimated values” of “resource estimation prediction” via the prediction model); and
for an identified absence of a change to the relationship between the energy usage and the associated characteristics beyond a threshold based on the second prediction error being within a range between a third value and a fourth value (Li, sections 4D, 6, and Algorithm 1 teach iteratively using a validation set for computing “validation error” (first prediction error associated with the second time period) of the model; wherein “Q outputs the predicted energy and temperature data, concatenated as yr”, and ensuring the minimum error is the same “between the predicted yr (relationship between the energy usage and the associated characteristics) and the real data” used in the error computation that is further compared to predetermined values, “EQval” (beyond a threshold based on the second prediction error being within a range between a third value and a fourth value)), wherein the third value and the fourth value are based on at least one of the additional average for the third prediction error or the additional standard deviation for the third prediction error:
using the first machine-trained model to predict the first predicted energy usage (Li, sections 4D, 6, and Algorithm 1 teach iteratively using a validation set for computing “validation error” of the model (first machine-trained model) that can be an MAE calculations compared to predetermined values (the third value and the fourth value are based on at least one of the additional average for the third prediction error); wherein “Q outputs the predicted energy and temperature data, concatenated as yr”, and ensuring).
Li, Stein, Baig, and Cambron are combinable for the same rationale as set forth above with respect to claim 9.
Regarding claim 12, the combination of Li and Stein teach all the claim limitations of claim 11 above; and further teach updating the first machine-trained model based on the fifth data and at least a subset of the first data, wherein the first predicted energy usage for the second time period is based on the first machine-trained model after the updating of the first machine-trained model based on the fifth data and at least the subset of the first data (Baig, sections 4, 6.1.2-6.3, and 7 teach determining “MSE” (additional average) for the different window training sets (first/fifth time period) for the “true and estimated values” of “resource estimation prediction” via the prediction model and used for training the model (updating of the first machine-trained model)).
Li, Stein, Baig, and Cambron are combinable for the same rationale as set forth above with respect to claim 11.
Regarding claim 20, claims 9-10 are analogous and the combination of Li, Stein, Baig, and Cambron teach all the claim limitations of claims 9-10 above.
Prior Art
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Schline et al (US Pub 20210135478) teaches utilizing EWMA for power calculations in a machine learning model applied to datacenter functions.
Conclusion
19. Any inquiry concerning this communication or earlier communications from the examiner should be directed to CLINT MULLINAX whose telephone number is 571-272-3241. The examiner can normally be reached on Mon - Fri 8:00-4:30 PT.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Alexey Shmatov can be reached on 571-270-3428. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of an application may be obtained from the Patent Application Information Retrieval (PAIR) system. Status information for published applications may be obtained from either Private PAIR or Public PAIR. Status information for unpublished applications is available through Private PAIR only. For more information about the PAIR system, see http://pair-direct.uspto.gov. Should you have questions on access to the Private PAIR system, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative or access to the automated information system, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/C.M./Examiner, Art Unit 2123
/ALEXEY SHMATOV/Supervisory Patent Examiner, Art Unit 2123
Prosecution Insights