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 .
Remarks
This Office Action is in response to applicant’s amendment filed on May 18, 2026, under which claims 1-20 are pending and under consideration.
Response to Arguments
Applicant’s amendments have overcome the previous rejections under § 101 and § 102. Therefore, these rejections have been withdrawn.
Applicant’s amendments have also overcome the previous rejections under § 103. However, upon further consideration, new grounds of rejection have been made under § 103 based on a new reference, Varamesh. Applicant’s arguments are moot under the new grounds of rejection, since they rely on new limitations that are addressed by the new reference.
Claim Interpretation
The following is a quotation of 35 U.S.C. 112(f):
(f) Element in Claim for a Combination. – An element in a claim for a combination may be expressed as a means or step for performing a specified function without the recital of structure, material, or acts in support thereof, and such claim shall be construed to cover the corresponding structure, material, or acts described in the specification and equivalents thereof.
The following is a quotation of pre-AIA 35 U.S.C. 112, sixth paragraph:
An element in a claim for a combination may be expressed as a means or step for performing a specified function without the recital of structure, material, or acts in support thereof, and such claim shall be construed to cover the corresponding structure, material, or acts described in the specification and equivalents thereof.
The claims in this application are given their broadest reasonable interpretation using the plain meaning of the claim language in light of the specification as it would be understood by one of ordinary skill in the art. The broadest reasonable interpretation of a claim element (also commonly referred to as a claim limitation) is limited by the description in the specification when 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, is invoked.
As explained in MPEP § 2181, subsection I, claim limitations that meet the following three-prong test will be interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph:
(A) the claim limitation uses the term “means” or “step” or a term used as a substitute for “means” that is a generic placeholder (also called a nonce term or a non-structural term having no specific structural meaning) for performing the claimed function;
(B) the term “means” or “step” or the generic placeholder is modified by functional language, typically, but not always linked by the transition word “for” (e.g., “means for”) or another linking word or phrase, such as “configured to” or “so that”; and
(C) the term “means” or “step” or the generic placeholder is not modified by sufficient structure, material, or acts for performing the claimed function.
Use of the word “means” (or “step”) in a claim with functional language creates a rebuttable presumption that the claim limitation is to be treated in accordance with 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph. The presumption that the claim limitation is interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, is rebutted when the claim limitation recites sufficient structure, material, or acts to entirely perform the recited function.
Absence of the word “means” (or “step”) in a claim creates a rebuttable presumption that the claim limitation is not to be treated in accordance with 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph. The presumption that the claim limitation is not interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, is rebutted when the claim limitation recites function without reciting sufficient structure, material or acts to entirely perform the recited function.
Claim limitations in this application that use the word “means” (or “step”) are being interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, except as otherwise indicated in an Office action. Conversely, claim limitations in this application that do not use the word “means” (or “step”) are not being interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, except as otherwise indicated in an Office action.
This application includes one or more claim limitations that do not use the word “means,” but are nonetheless being interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, because the claim limitation(s) uses a generic placeholder that is coupled with functional language without reciting sufficient structure to perform the recited function and the generic placeholder is not preceded by a structural modifier.
Claims 16-20 include limitations that invoke § 112(f). Such claim limitation(s) are:
“forecasting component configured to generate…” in claim 16
“training component configured to update…” in claim 17
“data collection component configured to collect…” in claim 19
“change attribution component configured to filter…” in claim 20
Here, “component” is a generic placeholder equivalent to “means” and “configured to” is a linking phrase equivalent to “for.” See MPEP § 2181.
Because this/these claim limitation(s) is/are being interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, it/they is/are being interpreted to cover the corresponding structure described in the specification as performing the claimed function, and equivalents thereof.
Support is found in paragraphs [0096], [0106], [0108], and [0110] of the specification, which teaches a general-purpose computer implementing the components using software.
It is also noted that claim 18 does not negate the § 112(f) interpretation of the forecasting component in claim 16, since claim 18 only generally associates the operation of the forecasting component with the processor and memory rather than specifically define the forecasting component as software executed by the recited processor. In other words, claim 18 does not exclude the forecasting component as being a device separate from the memory and processor.
If applicant does not intend to have this/these limitation(s) interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, applicant may: (1) amend the claim limitation(s) to avoid it/them being interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph (e.g., by reciting sufficient structure to perform the claimed function); or (2) present a sufficient showing that the claim limitation(s) recite(s) sufficient structure to perform the claimed function so as to avoid it/them being interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
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.
1. Claims 1-3, 5-6, 11-14, and 16-19 are rejected under 35 U.S.C. 103 as being unpatentable over Chen et al., “Outage Prediction and Diagnosis for Cloud Service Systems,” WWW '19: World Wide Web Conference (May 2019) (“Chen”) in view of Varamesh et al., “Mixture Dense Regression for Object Detection and Human Pose Estimation,” arXiv: 1912.00821v2 [cs.CV] 20 Apr 2020 (“Varmesh”) and Gugulothu et al. (US 2020/0401888 A1) (“Gugulothu”).
As to claim 1, Chen teaches a method for outage forecasting, comprising:
receiving, by a machine learning model, time series data for a service metric of a computer network; [§ 4.1.1, paragraph 2: “Given alerting signals across the cloud system, we predict whether a component or service outage will happen. For evaluation, we choose the service and component outages that occurred frequently in the last year…” The alerting signal is a time series, as defined in § 3.1: “A single alerting signal Ai can be represented by Ai …, where each component
A
t
i
indicates the strength of this signal at time t ∈ (1, T) …The input feature At …are the m alerting signals used for prediction.” The model is a machine learning model, which includes a “Bayesian network” and a “gradient boosting tree” as described in § 3.2 and § 3.3.]
[…] wherein the machine learning model comprises […] a second output head […] wherein the machine learning model is […] trained using […] a classification loss based on a classification output of the second output head; [§ 3.3, paragraph 2: “The XGBoost is optimized to achieve the best prediction results by minimizing the following loss… where the first term …is the summation of cross-entropy loss measuring whether the classification model and features can well perform for the prediction task across all timestamp.” Note that in this context, “optimized” refers to the training of the model (see also § 4.1.1, paragraph 1, which describes “training data” for the “training phase”).] and
generating, by a forecasting component, outage forecasting information for the computer network […]. [§ 3.3, paragraph 2: “In the CART tree, each node is one of the alerting signals. The prediction result is the sum of scores predicted by K decision trees…” § 4.1.1, paragraph 2: “Given alerting signals across the cloud system, we predict whether a component or service outage will happen.]
Chen does not teach:
(1) “generating, by the machine learning model, probability distribution information for the service metric based on the time series data”;
(2) wherein the machine learning model also comprises “a first output head” wherein the second output is “separate from the first output head”;
(3) the machine learning model being trained “jointly…using a distribution loss based on a distribution output of the first output head” and the classification loss; and
(4) the generation of the outage forecasting information being “based on the probability distribution information.”
Varamesh teaches “generating, by the machine learning model, probability distribution information” [§ 3.1, paragraph 2: “In the context of neural networks, mixture density networks (MDN) [4] enable us to use a neural network to estimate the parameters of a mixture model. An MDN estimates the probability density of a target vector conditioned on the input…. The objective of MDN is to fit the weights of a neural network such that it maximizes the likelihood of the training data.” Referring also to equations (3)-(4) and the text above and below these equations, the mixture density network computes the density function ϕm, based on the standard deviation σm and the mixing coefficient α (See text below equation (1) in § 3.1). That is, the probability in equation (1) and (3) are regarded as probability distribution information.] with a model comprising “a first output head” wherein a second output head is “separate from the first output head,” [FIG. 1, teaches the output Cp’ of the classification branch, which is a classification head analogous to the second output head, and the remaining branch that outputs (α, ϕ, σ), which correspond to a first output head separate from the second output head. This is described in § 3.2, paragraph 1: “The classification branch is realized by doing dense binary classification for each category y ∈Y. The number of categories for object detection is equal to the number of classes in a dataset, and for human pose estimation, it only includes the person class. Beside the classification branch, at each location we also need to predict pose parameters of interest T.” Note that the part that outputs the other outputs is called a “regression head” (see § 4.3.1). See also § 3.2, paragraph 3: “Given an input RGB image I of size H * W * 3, a CNN processes I and generates an output of dimensions H’ * W’ * C’. Here we have H = D * H’ (and similarly for width), where D is the down sampling factor of the network. We indicate the set of all output cells using P. At p 2 P, the output channels C0, includes Y binary classification channels. It also includes pose parameters T…” Since the classification branch and the regression head are separate outputs of the model, they are considered to be separate heads.], and the model being “jointly trained using a distribution loss based on a distribution output of the first output head” and a classification loss [§ 3.2, text below equation (5): “We do not modify the other loss terms used in Center-Net. This includes a binary classification loss LC for each class, offset regression loss term LCoff to compensate for lost spatial precision due to downsampling, and the term LT loss for pose parameters. The total loss is given in Eq. 6.” That is, the LT is a distribution loss, as described in § 3.2, text above equation (5) (“Given the conditional probability density of the ground truth in Eq. 3, we can define the regression objective as the negative log-likelihood and minimize it using stochastic gradient descent. In Eq. 5, we provide the negative loglikelihood for the pose targets generated by MDN, where N is the number of samples in the dataset.”) and is used in conjunction with the classification loss. Note that the loss here is used for training (see § 3.2, paragraph 2). Since the total loss includes both loss terms, the training of the model is performed “jointly” using those loss terms.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of Chen the teachings of Varamesh by implementing the technique of training and using a multi-head model of Varamesh, as taught in Varmaesh, so as to arrive at the above features of the instant claim. The motivation would have been to implement a model component that is capable of estimating the density of a distribution with the benefit of avoiding convergence to an average target given an input (see Varamesh § 3.1, paragraph 1: “Mixture models are powerful tools for estimating the density of any distribution [29]. Ideally, they recover modes that contribute to the generation of data and their distribution. For a regression task, mixture models help to avoid converging to an average target given an input.”) while maintaining the ability to perform classification (see Varamesh, § 3.2, paragraph 2: “However, spatial regression is a multi-modal task, and we believe modeling it using a single-mode network will lead to a suboptimal prediction.”).
The combination of references thus far does not explicitly teach the limitation of “for the service metric based on the time series data” for the step of generating the probability distribution information, and the limitation that the generation of the outage forecasting information is “based on the probability distribution information.”
Gugulothu, which generally relates to time series forecasting using machine learning models (see title and abstract), teaches generating “for the service metric based on the time series data,” [[0042]: “The outputs of the MDN (represented as 218 in FIG. 2A/258 in FIG. 2B) as formulated in (2) model the conditional distribution of the future values yt+1, …, t+p to be predicted given the latent representation zt expressed as follows.” [0043]: “Thus, the MDN layer outputs a well-defined joint probability distribution obtained for all the time steps in the forecast time horizon.” Note that the latent representation is based on a time series. See [0032]: “Given the input time series the encoder learns a latent representation zt of the time series.”], and the generation of forecasting information “based on the probability distribution information.” [See parts cited above. See also [0023]: “Finally, mixture density networks are used to model the trend shifts and variability present in the data and provide a confidence estimate of the prediction.” That is, the confidence estimate is a forecasting information.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of the references combined thus far with the teachings of Gugulothu by implementing the mixture density network (MDN) for confidence estimation for forecasting, so as to arrive at the claimed invention. The motivation would have been to use a type of model, suitable for time series forecasting in general, that enables modeling of variability and trend shifts present in the input and determination of a confidence estimate of a prediction, as suggested by Gugulothu (see, e.g., [0023]: “Finally, mixture density networks are used to model the trend shifts and variability present in the data and provide a confidence estimate of the prediction.”).
As to claim 2, the combination of Chen, Varamesh, and Gugulothu teaches the method of claim 1, further comprising:
identifying, by a data collection component, a plurality of service metrics; [Chen, § 4.1.1, paragraph 2: “Given alerting signals across the cloud system, we predict whether a component or service outage will happen.” As defined in Chen, § 3.1, A is a multivariate time series (“For all alerting signals, we denote a multivariate time series of length T as A… for all signals”), i.e., a plurality of service metrics. Note that examples of signals are shown in Chen, FIG. 2(a). In regards to the limitation of a “data collection component,” this is taught in the form of a software implementation of the functions stated in the instant claim. Chen implies that its method is implemented using software, since it involves complex operations on large datasets.]
computing, by the data collection component, correlation information for the plurality of service metrics; [Chen, § 3.2, paragraph 2: “In our method, the conditional dependence between the alerting signal and outage given a set of other alerting signals are obtained by calculating the Pearson correlation. The influence of the conditional signals needs to be regressed out first. Then, the Fisher’s z-transform is performed as in [5]. For example, given the time series sequence of an alerting signal Ai and an outage Oi, and the conditional set only contains the alerting signal Ai2, the correlation between Ai and Oi given Ai2 is… z is the correlation score for the alerting signal Ai and outage Oi give Ai2 . The significant test of z is then used to check whether the independence assumption can be accepted that Ai is independent of Oi given Ai2.” That is, the correlation score is a benchmark indicator that constitutes correlation information for the respective Ai, across the plurality of service metrics in A.] and
filtering, by the data collection component, the plurality of service metrics based on the correlation information to obtain the service metric. [Chen, § 3.2, last paragraph: “In our work, the Bayesian network mainly works as a diagnositic tool to infer the relationship between the alerting signals and the outage. We can also use it to select the most relevant features and feed them as the inputs of the outage prediction model.” Chen, § 4.1.2, paragraph 2: “This observation shows that Bayesian network can help us find the most representative and dependent alerting signals. Other signals which are not found to be highly related to outages in the network would not significantly improve the performance.” That is, the signals selected by the Bayesian network are considered to be “filtered” in that it is a sub selection of the original set.]
As to claim 3, the combination of Chen, Varamesh, and Gugulothu teaches the method of claim 1, further comprising:
identifying, by a data collection component, a plurality of service metrics; [Chen, § 4.1.1, paragraph 2: “Given alerting signals across the cloud system, we predict whether a component or service outage will happen.” As defined in Chen, § 3.1, A is a multivariate time series (“For all alerting signals, we denote a multivariate time series of length T as A… for all signals”), i.e., a plurality of service metrics. Note that examples of signals are shown in Chen, FIG. 2(a). In regards to the limitation of a “data collection component,” this is taught in the form of a software implementation of the functions stated in the instant claim. Chen implies that its method is implemented using software, since it involves complex operations on large datasets.]
identifying, by the data collection component, one or more benchmark indicators; [Chen, § 3.2, paragraph 2: “In our method, the conditional dependence between the alerting signal and outage given a set of other alerting signals are obtained by calculating the Pearson correlation. The influence of the conditional signals needs to be regressed out first. Then, the Fisher’s z-transform is performed as in [5]. For example, given the time series sequence of an alerting signal Ai and an outage Oi, and the conditional set only contains the alerting signal Ai2, the correlation between Ai and Oi given Ai2 is… z is the correlation score for the alerting signal Ai and outage Oi give Ai2 . The significant test of z is then used to check whether the independence assumption can be accepted that Ai is independent of Oi given Ai2.” That is, the correlation score is a benchmark indicator.] and
selecting, by the data collection component, the service metric from the plurality of service metrics based on the one or more benchmark indicators. [Chen, § 3.2, last paragraph: “In our work, the Bayesian network mainly works as a diagnositic tool to infer the relationship between the alerting signals and the outage. We can also use it to select the most relevant features and feed them as the inputs of the outage prediction model.” Chen, § 4.1.2, paragraph 2: “This observation shows that Bayesian network can help us find the most representative and dependent alerting signals. Other signals which are not found to be highly related to outages in the network would not significantly improve the performance.” That is, the signals selected by the Bayesian network are considered to be “filtered” in that it is a sub selection of the original set.]
As to claim 5, the combination of Chen, Varamesh, and Gugulothu teaches the method of claim 1, as set forth above.
Gugulothu further teaches further comprising:
encoding the time series data using a recurrent neural network to obtain encoded data; [[0025]: “In an embodiment, the sparse RMDN may include long-short term memory (LSTM) or encoder-decoder (ED) as the underlying recurrent architectures.” See also [0026]. [0032]: “An ED is a seq2seq learning model that contains a pair of RNNs (called encoder and decoder) which are trained simultaneously. Given the input time series the encoder learns a latent representation zt of the time series.”] and
decoding the encoded data using a mixture density network to obtain mixture parameters for a plurality of distributions, [[0032]: “The decoder, which has the same structure as the encoder, decodes the hidden state zt to predict y′t+1, . . . , t+p.” The decoder is part of a mixture density network. See [0002]: “time series prediction using a sparse recurrent mixture density network (RMDN), such as sparse LSTM-MDN and a sparse ED-MDN, for accurate forecasting of a high variability time series.”; “The outputs of the MDN (represented as 218 in FIG. 2A/258 in FIG. 2B) as formulated in (2) model the conditional distribution of the future values yt+1, …, t+p to be predicted given the latent representation zt expressed as follows.”] wherein the probability distribution information is based on the mixture parameters. [[0036]: “Every forecasted point of the time series is associated with its own mixture of Gaussians. Let K be the total number of mixtures, then each component k∈{1, . . . , K} in the mixture is associated with coefficient ρk, mean μk and standard deviation σk.” [0047]: “To get prediction at time t, a Gaussian mixture k with the one having highest value of probability ρt,k is selected at 218 in FIG. 2A and 258 in FIG. 2B. The selected Gaussian mixture's mean μt,k is selected as the prediction and the standard deviation σt,k confidence estimate of the prediction at 218 in FIG. 2A and 258 in FIG. 2B.”]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of the references combined thus far with the above teachings of Gugulothu to have also arrived at the limitations of the instant claim. The motivation for doing so is the same as the one given for the teachings of Gugulothu in the rejection of the parent claim, since the techniques of Gugulothu discussed for the instant dependent claim are part of those discussed in the rejection of the parent independent claim.
As to claim 6, the combination of Chen, Varamesh, and Gugulothu teaches the method of claim 5, as set forth above.
Gugulothu further teaches further comprising:
generating a mixing coefficient using the mixture density network, wherein the probability distribution information is based on the mixing coefficient. [[0036]: “Every forecasted point of the time series is associated with its own mixture of Gaussians. Let K be the total number of mixtures, then each component k∈{1, . . . , K} in the mixture is associated with coefficient ρk, mean μk and standard deviation σk.”]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of the references combined thus far with the above teachings of Gugulothu to have also arrived at the limitations of the instant claim. The motivation for doing so is the same as the one given for the teachings of Gugulothu in the rejection of the parent claim, since the techniques of Gugulothu discussed for the instant dependent claim are part of those discussed in the rejection of the parent independent claim.
As to claim 11, Chen teaches a method for training a machine learning model, comprising:
receiving, by a training component, training data including time series data for a service metric of a computer network and outage data for the computer network; [§ 4.1.1, paragraph 1: “We acquired data at the time step of one hour and obtained over 8,000 samples in total (24hrs*365days). …To better deal with the imbalanced data, the SMOTE [2] over-sampling strategy is used for generating the training data from system database so that the positive and negative samples in the training phase can be balanced.”] […]
generating, by a second output head of the machine learning model […], threshold outage information based on the time series data; [Chen, § 3.2, paragraphs 2-3: “In our method, the conditional dependence between the alerting signal and outage given a set of other alerting signals are obtained by calculating the Pearson correlation. The influence of the conditional signals needs to be regressed out first. Then, the Fisher’s z-transform is performed as in [5]. For example, given the time series sequence of an alerting signal Ai and an outage Oi, and the conditional set only contains the alerting signal Ai2, the correlation between Ai and Oi given Ai2 is… z is the correlation score for the alerting signal Ai and outage Oi give Ai2 . The significant test of z is then used to check whether the independence assumption can be accepted that Ai is independent of Oi given Ai2…In our work, the Bayesian network mainly works as a diagnositic tool to infer the relationship between the alerting signals and the outage. We can also use it to select the most relevant features and feed them as the inputs of the outage prediction model.” That is, the correlation score is used as information to determine whether a particular signal meets a relevance threshold.] and
[…] training, by the training component, parameters of the machine learning model based on […] and a classification loss […], and wherein the classification loss is based on the threshold outage information, and the outage data. [Chen, § 3.3, paragraph 2: “The XGBoost is optimized to achieve the best prediction results by minimizing the following loss… where the first term …is the summation of cross-entropy loss measuring whether the classification model and features can well perform for the prediction task across all timestamp.” Note that in this context, “optimized” refers to the training of the model (see also § 4.1.1, paragraph 1, which describes “training data” for the “training phase”).]
Chen does not teach:
(1) “generating, by a first output head of a machine learning model, probability distribution information for the service metric based on the time series data”;
(2) the machine learning model also comprises “a first output head” wherein the second output head is “separate from the first output head”; and
(3) the training being performed “jointly” and “based on a distribution loss” in addition to the classification loss, “wherein the distribution loss is based on the probability distribution information.”
Varamesh teaches “generating, by a first output head of the machine learning model, probability distribution information” [§ 3.1, paragraph 2: “In the context of neural networks, mixture density networks (MDN) [4] enable us to use a neural network to estimate the parameters of a mixture model. An MDN estimates the probability density of a target vector conditioned on the input…. The objective of MDN is to fit the weights of a neural network such that it maximizes the likelihood of the training data.” Referring also to equations (3)-(4) and the text above and below these equations, the mixture density network computes the density function ϕm, based on the standard deviation σm and the mixing coefficient α (See text below equation (1) in § 3.1). That is, the probability in equation (1) and (3) are regarded as probability distribution information.] and “a first output head” wherein the second output head is “separate from the first output head” [In regards to “a first output head” wherein a second output head is “separate from the first output head,” FIG. 1, teaches the output Cp’ of the classification branch, which is a classification head analogous to the second output head, and the remaining branch that outputs (α, ϕ, σ), which correspond to a first output head separate from the second output head. This is described in § 3.2, paragraph 1: “The classification branch is realized by doing dense binary classification for each category y ∈Y. The number of categories for object detection is equal to the number of classes in a dataset, and for human pose estimation, it only includes the person class. Beside the classification branch, at each location we also need to predict pose parameters of interest T.” Note that the part that outputs the other outputs is called a “regression head” (see § 4.3.1). See also § 3.2, paragraph 3: “Given an input RGB image I of size H * W * 3, a CNN processes I and generates an output of dimensions H’ * W’ * C’. Here we have H = D * H’ (and similarly for width), where D is the down sampling factor of the network. We indicate the set of all output cells using P. At p 2 P, the output channels C0, includes Y binary classification channels. It also includes pose parameters T…” Since the classification branch and the regression head are separate outputs of the model, they are considered to be separate heads.] and the model being “jointly” “based on a distribution loss” in addition to a classification loss “wherein the distribution loss is based on the probability distribution information.” [§ 3.2, text below equation (5): “We do not modify the other loss terms used in Center-Net. This includes a binary classification loss LC for each class, offset regression loss term LCoff to compensate for lost spatial precision due to downsampling, and the term LT loss for pose parameters. The total loss is given in Eq. 6.” That is, the LT is a distribution loss, as described in § 3.2, text above equation (5) (“Given the conditional probability density of the ground truth in Eq. 3, we can define the regression objective as the negative log-likelihood and minimize it using stochastic gradient descent. In Eq. 5, we provide the negative loglikelihood for the pose targets generated by MDN, where N is the number of samples in the dataset.”) and is used in conjunction with the classification loss. Note that the loss here is used for training (see § 3.2, paragraph 2). Since the total loss includes both loss terms, the training of the model is performed “jointly” using those loss terms.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of Chen the teachings of Varamesh by implementing the technique of training and using a multi-head model of Varamesh, as taught in Varmaesh, so as to arrive at the above features of the instant claim. The motivation would have been to implement a model component that is capable of estimating the density of a distribution with the benefit of avoiding convergence to an average target given an input (see Varamesh § 3.1, paragraph 1: “Mixture models are powerful tools for estimating the density of any distribution [29]. Ideally, they recover modes that contribute to the generation of data and their distribution. For a regression task, mixture models help to avoid converging to an average target given an input.”) while maintaining the ability to perform classification (see Varamesh, § 3.2, paragraph 2: “However, spatial regression is a multi-modal task, and we believe modeling it using a single-mode network will lead to a suboptimal prediction.”).
The combination of references thus far does not explicitly teach the limitation of “for the service metric based on the time series data” for the step of generating the probability distribution information.
Gugulothu, which generally relates to time series forecasting using machine learning models (see title and abstract), teaches generating “for the service metric based on the time series data,” [[0042]: “The outputs of the MDN (represented as 218 in FIG. 2A/258 in FIG. 2B) as formulated in (2) model the conditional distribution of the future values yt+1, …, t+p to be predicted given the latent representation zt expressed as follows.” [0043]: “Thus, the MDN layer outputs a well-defined joint probability distribution obtained for all the time steps in the forecast time horizon.” Note that the latent representation is based on a time series. See [0032]: “Given the input time series the encoder learns a latent representation zt of the time series.”], and the generation of the forecasting information “based on the probability distribution information.” [See parts cited above. See also [0023]: “Finally, mixture density networks are used to model the trend shifts and variability present in the data and provide a confidence estimate of the prediction.” That is, the confidence estimate is a forecasting information.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of the references combined thus far with the teachings of Gugulothu by implementing the mixture density network (MDN) for confidence estimation for forecasting, so as to arrive at the claimed invention. The motivation would have been to use a type of model, suitable for time series forecasting in general, that enables modeling of variability and trend shifts present in the input and determination of a confidence estimate of a prediction, as suggested by Gugulothu (see, e.g., [0023]: “Finally, mixture density networks are used to model the trend shifts and variability present in the data and provide a confidence estimate of the prediction.”).
As to claim 12, combination of Chen, Varamesh, and Gugulothu teaches the method of claim 11, as set forth above.
Gugulothu further teaches further comprising:
encoding, by a recurrent neural network of the machine learning model, the time series data to obtain encoded data; [[0025]: “In an embodiment, the sparse RMDN may include long-short term memory (LSTM) or encoder-decoder (ED) as the underlying recurrent architectures.” See also [0026]. [0032]: “An ED is a seq2seq learning model that contains a pair of RNNs (called encoder and decoder) which are trained simultaneously. Given the input time series the encoder learns a latent representation zt of the time series.”] and
decoding, by a mixture density network of the machine learning model, the encoded data to obtain mixture parameters for a plurality of distributions, [[0032]: “The decoder, which has the same structure as the encoder, decodes the hidden state zt to predict y′t+1, . . . , t+p.” The decoder is part of a mixture density network. See [0002]: “time series prediction using a sparse recurrent mixture density network (RMDN), such as sparse LSTM-MDN and a sparse ED-MDN, for accurate forecasting of a high variability time series.”; “The outputs of the MDN (represented as 218 in FIG. 2A/258 in FIG. 2B) as formulated in (2) model the conditional distribution of the future values yt+1, …, t+p to be predicted given the latent representation zt expressed as follows.”] wherein the probability distribution information is based on the mixture parameters. [[0036]: “Every forecasted point of the time series is associated with its own mixture of Gaussians. Let K be the total number of mixtures, then each component k∈{1, . . . , K} in the mixture is associated with coefficient ρk, mean μk and standard deviation σk.” [0047]: “To get prediction at time t, a Gaussian mixture k with the one having highest value of probability ρt,k is selected at 218 in FIG. 2A and 258 in FIG. 2B. The selected Gaussian mixture's mean μt,k is selected as the prediction and the standard deviation σt,k confidence estimate of the prediction at 218 in FIG. 2A and 258 in FIG. 2B.”]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of the references combined thus far with the above teachings of Gugulothu to have also arrived at the limitations of the instant claim. The motivation for doing so is the same as the one given for the teachings of Gugulothu in the rejection of the parent claim, since the techniques of Gugulothu discussed for the instant dependent claim are part of those discussed in the rejection of the parent independent claim.
As to claim 13, the combination of Chen, Varamesh, and Gugulothu teaches the method of claim 11, as set forth above.
Gugulothu further teaches “further comprising: computing, by the training component, a distribution loss based on the probability distribution information and the outage data, wherein the parameters of the machine learning model are updated based on the distribution loss.” [[0043]: “The model parameters are learned by minimizing the negative log-likelihood of the distribution in (3) as shown below.” Note that “learned by minimizing” refers to the training process of the model.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of the references combined thus far with the above teachings of Gugulothu to have also arrived at the limitations of the instant claim. The motivation for doing so is the same as the one given for the teachings of Gugulothu in the rejection of the parent claim, since the techniques of Gugulothu discussed for the instant dependent claim are part of those discussed in the rejection of the parent independent claim.
As to claim 14, the combination of Chen, Varamesh, and Gugulothu teaches the method of claim 11, further comprising:
computing, by the training component, a binary cross-entropy loss based on the threshold outage information and outage label data, wherein the parameters of the machine learning model are updated based on the binary cross-entropy loss. [Chen, § 3.3, paragraph 2: “The XGBoost is optimized to achieve the best prediction results by minimizing the following loss… where the first term …is the summation of cross-entropy loss measuring whether the classification model and features can well perform for the prediction task across all timestamp.” Note that in this context, “optimized” refers to the training of the model (see also § 4.1.1, paragraph 1, which describes “training data” for the “training phase”). The limitation of “outage label data” is taught in § 3.1, paragraph 2: “The outage sequence is a binary time series of length T as O = (O1,O2, ...,OT ) ∈ RT , where for each t ∈ {1, 2, ...,T }, Ot ∈ {1, 0} indicating whether an outage happens at time t or not.”]
As to claim 16, Chen teaches an apparatus for outage forecasting, [Chen teaches that its method is implemented on a computer apparatus. See, e.g., § 2, paragraph 2: “When an outage is detected, the management tool is expected to automatically notify, mitigate, and diagnose the outage…our work uses the failure signals from monitors (which we call, alerting signals) to predict critical failures (which we call, outages). In current real-time system, outage prediction is an important issue that bothers many systems.”] comprising:
a machine learning model configured to generate […] for a service metric of a computer network based on time series data, wherein the machine learning model comprises a […] second output head […] [§ 4.1.1, paragraph 2: “Given alerting signals across the cloud system, we predict whether a component or service outage will happen. For evaluation, we choose the service and component outages that occurred frequently in the last year…” The alerting signal is a time series, as defined in § 3.1: “A single alerting signal Ai can be represented by Ai …, where each component
A
t
i
indicates the strength of this signal at time t ∈ (1, T) …The input feature At …are the m alerting signals used for prediction.” The model is a machine learning model, which includes a “Bayesian network” and a “gradient boosting tree” as described in § 3.2 and § 3.3.] wherein the machine learning model is […] trained using […] and a classification loss based on a classification output of the second output head; [§ 3.3, paragraph 2: “The XGBoost is optimized to achieve the best prediction results by minimizing the following loss… where the first term …is the summation of cross-entropy loss measuring whether the classification model and features can well perform for the prediction task across all timestamp.” Note that in this context, “optimized” refers to the training of the model (see also § 4.1.1, paragraph 1, which describes “training data” for the “training phase”).] and
a forecasting component configured to generate outage forecasting information for the computer network […]. [§ 3.3, paragraph 2: “In the CART tree, each node is one of the alerting signals. The prediction result is the sum of scores predicted by K decision trees…” § 4.1.1, paragraph 2: “Given alerting signals across the cloud system, we predict whether a component or service outage will happen.]
Chen does not teach:
(1) the machine learning model also comprises “a first output head” wherein the second output head is “separate from the first output head”;
(2) the model being trained “jointly…using a distribution loss based on a distribution output of the first output head” in addition to using the classification loss; and
(3) the generation of the outage forecasting information being “based on the probability distribution information.”
Varamesh teaches “a first output head of the machine learning model, probability distribution information” [§ 3.1, paragraph 2: “In the context of neural networks, mixture density networks (MDN) [4] enable us to use a neural network to estimate the parameters of a mixture model. An MDN estimates the probability density of a target vector conditioned on the input…. The objective of MDN is to fit the weights of a neural network such that it maximizes the likelihood of the training data.” Referring also to equations (3)-(4) and the text above and below these equations, the mixture density network computes the density function ϕm, based on the standard deviation σm and the mixing coefficient α (See text below equation (1) in § 3.1). That is, the probability in equation (1) and (3) are regarded as probability distribution information.] and “a first output head” wherein the second output head is “separate from the first output head” [In regards to “a first output head” wherein a second output head is “separate from the first output head,” FIG. 1, teaches the output Cp’ of the classification branch, which is a classification head analogous to the second output head, and the remaining branch that outputs (α, ϕ, σ), which correspond to a first output head separate from the second output head. This is described in § 3.2, paragraph 1: “The classification branch is realized by doing dense binary classification for each category y ∈Y. The number of categories for object detection is equal to the number of classes in a dataset, and for human pose estimation, it only includes the person class. Beside the classification branch, at each location we also need to predict pose parameters of interest T.” Note that the part that outputs the other outputs is called a “regression head” (see § 4.3.1). See also § 3.2, paragraph 3: “Given an input RGB image I of size H * W * 3, a CNN processes I and generates an output of dimensions H’ * W’ * C’. Here we have H = D * H’ (and similarly for width), where D is the down sampling factor of the network. We indicate the set of all output cells using P. At p 2 P, the output channels C0, includes Y binary classification channels. It also includes pose parameters T…” Since the classification branch and the regression head are separate outputs of the model, they are considered to be separate heads.] and the model being trained “jointly” trained “using a distribution loss based on a distribution output of the first output head” and a classification loss [§ 3.2, text below equation (5): “We do not modify the other loss terms used in Center-Net. This includes a binary classification loss LC for each class, offset regression loss term LCoff to compensate for lost spatial precision due to downsampling, and the term LT loss for pose parameters. The total loss is given in Eq. 6.” That is, the LT is a distribution loss, as described in § 3.2, text above equation (5) (“Given the conditional probability density of the ground truth in Eq. 3, we can define the regression objective as the negative log-likelihood and minimize it using stochastic gradient descent. In Eq. 5, we provide the negative loglikelihood for the pose targets generated by MDN, where N is the number of samples in the dataset.”) and is used in conjunction with the classification loss. Note that the loss here is used for training (see § 3.2, paragraph 2). Since the total loss includes both loss terms, the training of the model is performed “jointly” using those loss terms.]
The combination of references thus far does not teach the limitation of the generation of the outage forecasting information being “based on the probability distribution information.”
Gugulothu, which generally relates to time series forecasting using machine learning models (see title and abstract), teaches generation of forecasting information “based on the probability distribution information.” [[0042]: “The outputs of the MDN (represented as 218 in FIG. 2A/258 in FIG. 2B) as formulated in (2) model the conditional distribution of the future values yt+1, …, t+p to be predicted given the latent representation zt expressed as follows.” [0043]: “Thus, the MDN layer outputs a well-defined joint probability distribution obtained for all the time steps in the forecast time horizon.” Note that the latent representation is based on a time series. See [0032]: “Given the input time series the encoder learns a latent representation zt of the time series.” See also [0023]: “Finally, mixture density networks are used to model the trend shifts and variability present in the data and provide a confidence estimate of the prediction.” That is, the confidence estimate is a forecasting information.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of Chen with the teachings of Gugulothu by implementing the training and the use of a mixture density network (MDN) for confidence estimation, as taught in Gugulothu, for the outage forecasting application taught in Chen, so as to arrive at the claimed invention. The motivation would have been to use a type of model, suitable for time series forecasting in general, that enables modeling of variability and trend shifts present in the input and determination of a confidence estimate of a prediction, as suggested by Gugulothu (see, e.g., [0023]: “Finally, mixture density networks are used to model the trend shifts and variability present in the data and provide a confidence estimate of the prediction.”).
As to claim 17, the combination of Chen, Varamesh, and Gugulothu teaches the apparatus of claim 16, further comprising:
a training component configured to update parameters of the machine learning model. [Chen, § 3.3, paragraph 2: “The XGBoost is optimized to achieve the best prediction results by minimizing the following loss… where the first term …is the summation of cross-entropy loss measuring whether the classification model and features can well perform for the prediction task across all timestamp.” Note that in this context, “optimized” refers to the training of the model (see also § 4.1.1, paragraph 1, which describes “training data” for the “training phase”).]
As to claim 18, the combination of Chen, Varamesh, and Gugulothu teaches the apparatus of claim 16, further comprising: a memory; and a processor configured to cause the machine learning model and the forecasting component to operate based on instructions stored in the memory. [As noted in the rejection of the parent claim, Chen teaches that its method is implemented on a computer system. Since the instant limitations are generic components of a general-purpose computer, the instant limitations are implied by Chen.]
As to claim 19, the combination Chen, Varamesh, and Gugulothu teaches the apparatus of claim 16, further comprising:
a data collection component configured to collect the time series data for the service metric. [Chen, § 4.1.1, paragraph 2: “Given alerting signals across the cloud system, we predict whether a component or service outage will happen.” As defined in Chen, § 3.1, A is a multivariate time series (“For all alerting signals, we denote a multivariate time series of length T as A… for all signals”), i.e., a plurality of service metrics. Note that examples of signals are shown in Chen, FIG. 2(a). In regards to the limitation of a “data collection component,” this is taught in the form of a software implementation of the functions stated in the instant claim. Chen implies that its method is implemented using software, since it involves complex operations on large datasets.]
2. Claim 4 is rejected under 35 U.S.C. 103 as being unpatentable over Chen in view of Varamesh and Gugulothu, and further in view of Wang et al. (US 2012/0136909 A1) (“Wang”).
As to claim 4, the combination of Chen, Varamesh, and Gugulothu teaches the method of claim 1, but does not teach the further limitations of the instant dependent claim.
Wang teaches further comprising:
collecting, by a data collection component, data for the service metric for a plurality of instances; [[0025]: “FIG. 4 is a diagram of an example EbAT framework 400. Shown are the data center 108 and various example metrics generated by this data center 108. These example metrics include service level metrics 401, system level metrics 402, and platform metrics 403. These metrics are provided to an EbAT enabled device or series of devices.”] and
computing, by the data collection component, an aggregate value for the service metric over the plurality of instances at each of a plurality of time steps, wherein the time series data is based on the aggregate value. [[0026]: “Operatively connected to the processor 501 is an aggregation module 506 to generate an entropy time series through transforming the at least one vector value into an entropy value to be displayed as part of a look-back window.” [0031]: “Operation 1005 is executed to generate an aggregated entropy time series for the parent through transforming all of the vector values for all of the children into the aggregated entropy time series.” [0020]: “The resulting entropy distributions aggregate raw metric data across the utility cloud to form “entropy time series,” and in addition, each hierarchy of the cloud (e.g., H-Crossing components such as data center, container, rack, enclosure, node, socket, core) can generate higher level from lower level entropy time series.”]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of the references combined thus far with the teachings of Wang by implementing the aggregation technique so as to arrive at the claimed invention of the instant claim. The motivation for doing so would have been to analyze entropy distributions aggregate raw metric data across a cloud, which is a form of information that is indicative of anomalies (see Wang, [0020]: “The resulting entropy distributions aggregate raw metric data across the utility cloud to form “entropy time series,”… Tools, such as spike detecting (e.g., visually or using time series analysis), signal processing or subspace method, are used to identify anomalies in entropy time series in general and at each level of the hierarchy.”).
3. Claims 7-8 are rejected under 35 U.S.C. 103 as being unpatentable over Chen in view of Varamesh and Gugulothu, and further in view of Yang et al. (US 2022/0382856 A1) (“Yang”).
As to claim 7, the combination of Chen, Varamesh, and Gugulothu teaches the method of claim 5, further comprising:
decoding the encoded data [As noted above in the rejection of claim 5, the step of decoding is taught as part of the model structure of Gugulothu incorporated in the combined teachings of the references.]
The combination of references thus far does not teach “using a classification network to obtain threshold outage information, wherein the outage forecasting information is based on the threshold outage information.”
Yang teaches “using a classification network to obtain threshold outage information, wherein the outage forecasting information is based on the threshold outage information.” [[0046]: “A point is labeled as an anomaly if its anomaly score is larger than a certain threshold, and thus the anomaly output 255 can be generated.” Since the anomaly score is compared to a threshold, it is considered to be a threshold information (note that the context of “outage” is already taught by the existing references), and the label is analogous to information that is based on threshold outage information.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of the references combined thus far with the teachings of Yang by implementing the model of Chen, as modified by Gugulothu, to include a classification network so as to arrive at the claimed invention. The motivation for doing so would have been to determine a causal graph from a multivariate time series input, which can be used to train a model for time series analysis (see Yang (see [0018]: “…Specifically, a causal graph is determined from a multivariate time series input, which describes the causal relationship between a plurality of causal variables in the multivariate time series input… The causal structure can then be served as a condition for a machine learning model to predict a conditional distribution of the multivariate time series, which is used to train the machine learning model.”), particularly in a manner that decompose the training complexity, which improves computational efficiency in training for real world applications (see Yang, [0019]: “The decomposability of the anomaly detection problem also helps to decompose the training complexity, which improves computational efficiency in training for real world applications and in particular for root cause analysis.”).
As to claim 8, the combination of Chen, Varamesh, and Gugulothu teaches the method of claim 1, but does not teach the further limitations of the instant dependent claim.
Yang teaches further comprising:
generating, by a change attribution component, a causal graph based on the time series data for the service metric and deployment data related to the service metric; [[0018]: “Specifically, a causal graph is determined from a multivariate time series input, which describes the causal relationship between a plurality of causal variables in the multivariate time series input. A subset of causal variables are than derived as having no causal parents in the causal graph. A causal structure of the multivariate time series can then be determined based on the subset of causal variables.” Note that the “multivariate time series input” is analogous to both the service metric and deployment data, since deployment data can also be another service metric that is relate to the first service metric. See also [0050]: “At step 302, the system may receive, via a data interface, training data containing a multivariate time series input (e.g., the multivariate time series data 230 in FIG. 2A) over a period of time.” [0016]: “For example, some existing systems may treat each performance metric individually using univariate time series anomaly detection algorithms, or alternatively treat all the performance metrics as an entity using multivariate time series anomaly detection algorithms.”] and
aggregating, by the change attribution component, causality information for a service based on the causal graph, wherein the outage forecasting information is based on the aggregated causality information. [[0018]: “Specifically, a causal graph is determined from a multivariate time series input, which describes the causal relationship between a plurality of causal variables in the multivariate time series input. A subset of causal variables are than derived as having no causal parents in the causal graph. A causal structure of the multivariate time series can then be determined based on the subset of causal variables. The causal structure can then be served as a condition for a machine learning model to predict a conditional distribution of the multivariate time series, which is used to train the machine learning model.” Here, the causal structure is regarded as an aggregation of causality information for the service. Regarding the limitation of “wherein the outage forecasting information is based on the aggregated causality information,” the training of a model used for forecasting would result in the predictions of the model being based on the causal structure. Moreover, the term “based on” does not require a specific relationship.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of the references combined thus far with the teachings of Yang by implementing the model of Chen, as modified by Gugulothu, to include a classification network so as to arrive at the claimed invention. The motivation for doing so would have been to determine a causal graph from a multivariate time series input, which can be used to train a model for time series analysis (see Yang (see [0018]: “…Specifically, a causal graph is determined from a multivariate time series input, which describes the causal relationship between a plurality of causal variables in the multivariate time series input… The causal structure can then be served as a condition for a machine learning model to predict a conditional distribution of the multivariate time series, which is used to train the machine learning model.”), particularly in a manner that decompose the training complexity, which improves computational efficiency in training for real world applications (see Yang, [0019]: “The decomposability of the anomaly detection problem also helps to decompose the training complexity, which improves computational efficiency in training for real world applications and in particular for root cause analysis.”).
4. Claim 9 is are rejected under 35 U.S.C. 103 as being unpatentable over Chen in view of Varamesh, Gugulothu and Yang, and further in view of Fattu et al. (US 2020/0162342 A1) (“Fattu”).
As to claim 9, the combination of Chen, Varamesh, Gugulothu, and Yang teaches the method of claim 8, as set forth above.
Fattu teaches “further comprising: filtering, by the change attribution component, a set of service changes corresponding to the deployment data based on the aggregated causality information.” [[0004]: “The preventative action module communicates a ranked set of recommended actions for mitigating a predicted failure of one or more of the replaceable devices.” [0095]: “For example, in response to a recommended action to replace one of the replaceable devices 118, the visual map displays a symbol with an easily visible color such as red, that illustrates the location or coordinates of replaceable device 118 which is to be replaced.” Replacement of a device corresponds to a service change, and ranking set of recommended actions for such constitutes an act of filtering. Additionally, the concept of being based on a predicted failure is analogous to “based on the aggregated causality information” when applied to the existing combination of references, since a recommended action based on a prediction is also based on causality information used to train the model. The limitation of “corresponding to deployment data” is met in the combination of references because Chen already teaches the service metric, as discussed above, and the data corresponding to a service metric can also correspond to deployment data.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of the references combined thus far with the teachings of Fattu by implementing its technique of determining recommendations actions in response to a predicted failure, so as to arrive at the limitations of the instant dependent claim. Doing so would enable mitigation of a predicted failure, as suggested by Fattu ([0004]: “A preventative action module is also included in the apparatus. The preventative action module communicates a ranked set of recommended actions for mitigating a predicted failure of one or more of the replaceable devices, the failure predicted based on a comparison between current weighted performance and environmental measurements and the baseline correlations”).
5. Claim 10 is rejected under 35 U.S.C. 103 as being unpatentable over Chen in view of Varamesh and Gugulothu, and further in view of Mehta et al. (US 2020/0184355 A1).
As to claim 10, the combination of Chen, Varamesh, and Gugulothu teaches the method of claim 1, but does not teach the further limitations of the instant dependent claim.
Mehta teaches “further comprising: determining, by the forecasting component, that a likelihood of an outage in the computer network exceeds a threshold based on the probability distribution information; and transmitting, by the forecasting component, an alert based on the determination.” [[0073]: “An alert is generated at the dashboard and/or disseminated via the network 120 when an outage prediction score exceeds a predefined prediction threshold. The prediction threshold may be customized by the application owner such that an alert is generated when the probability of an outage reaches an unacceptable level. The alert is intended to inform the application owner that the probability of an outage based on the input log data has reached the unacceptable level, thereby affording the application owner an opportunity to take prophylactic measures.” Note that the score is a classification of “likelihood” (probability). See [0069]: “As a result of applying the classifier algorithm to the several data inputs, the incident prediction engine 360 generates incident prediction scores classifying the likelihood of any known incidents occurring.” Therefore, the basis for the score as disclosed here is analogous to the probability distribution information of the instant claim.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of the references combined thus far with the teachings of Mehta by implementing the alerting technique as taught by Mehta so as to arrive at the limitations of the instant dependent claim. The motivation for doing so would have enabled “real-time diagnostic monitoring as well as alerts and outage counter-measures” (Mehta, [0065]: “The process flow involves inputting various data, which collectively make up decision criteria, to the incident prediction engine 360, and, based on the input data, providing real-time diagnostic monitoring as well as alerts and outage counter-measures.”).
6. Claim 15 is rejected under 35 U.S.C. 103 as being unpatentable over Chen in view of Varamesh and Gugulothu, and further in view of Zhang et al., “Modeling Extreme Events in Time Series Prediction,” KDD ’19, August 4–8, 2019, Anchorage, AK, USA (“Zhang”).
As to claim 15, the combination of Chen, Varamesh, and Gugulothu teaches the method of claim 11, further comprising:
computing, by the training component, an […] loss based on the threshold outage information and outage label data, wherein the parameters of the machine learning model are updated based on the extreme value loss. [Chen, § 3.3, paragraph 2: “The XGBoost is optimized to achieve the best prediction results by minimizing the following loss… where the first term …is the summation of cross-entropy loss measuring whether the classification model and features can well perform for the prediction task across all timestamp.” Note that in this context, “optimized” refers to the training of the model (see also § 4.1.1, paragraph 1, which describes “training data” for the “training phase”). The limitation of “outage label data” is taught in § 3.1, paragraph 2: “The outage sequence is a binary time series of length T as O = (O1,O2, ...,OT ) ∈ RT , where for each t ∈ {1, 2, ...,T }, Ot ∈ {1, 0} indicating whether an outage happens at time t or not.”]
The combination of references thus far does not teach the limitation of the loss including an “extreme value loss.”
Zhang teaches an “extreme value loss” [Abstract: “To address this issue, we take inspirations from the Extreme Value Theory, developing a new form of loss called Extreme Value Loss (EVL) for detecting the future occurrence of extreme events.”]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of the references combined thus far with the teachings of Zhang by implementing an extreme value loss so as to arrive at the limitations of the instant claim. The motivation would have been to implement a form of loss that enables time series prediction with extreme events (see Zhang, abstract: “By incorporating EVL with an adapted memory network module, we achieve an end-to-end framework for time series prediction with extreme events”).
7. Claim 20 is are rejected under 35 U.S.C. 103 as being unpatentable over Chen in view of Varamesh and Gugulothu, and further in view of Fattu et al. (US 2020/0162342 A1) (“Fattu”).
As to claim 20, the combination of Chen, Varamesh, and Gugulothu teaches the apparatus of claim 16, but does not teach the further limitations of the instant dependent claim.
Fattu teaches “further comprising: a change attribution component configured to filter a set of service changes corresponding to deployment data based on the probability distribution information.” [[0004]: “The preventative action module communicates a ranked set of recommended actions for mitigating a predicted failure of one or more of the replaceable devices.” [0095]: “For example, in response to a recommended action to replace one of the replaceable devices 118, the visual map displays a symbol with an easily visible color such as red, that illustrates the location or coordinates of replaceable device 118 which is to be replaced.” Replacement of a device corresponds to a service change, and ranking set of recommended actions for such constitutes an act of filtering. Additionally, the concept of being based on a predicted failure is analogous to “based on the probability distribution information” when applied to the existing combination of references. The limitation of “corresponding to deployment data” is met in the combination of references because Chen already teaches the service metric, as discussed above.]
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have combined the teachings of the references combined thus far with the teachings of Fattu by implementing its technique of determining recommendations actions in response to a predicted failure, so as to arrive at the limitations of the instant dependent claim. Doing so would enable mitigation of a predicted failure, as suggested by Fattu ([0004]: “A preventative action module is also included in the apparatus. The preventative action module communicates a ranked set of recommended actions for mitigating a predicted failure of one or more of the replaceable devices, the failure predicted based on a comparison between current weighted performance and environmental measurements and the baseline correlations”).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. The following documents depict the state of the art.
Yang et al., “Historical Context-based Style Classification of Painting Images via Label Distribution Learning,” MM’18, October 22-26, 2018, Seoul, Republic of Korea teaches the use of both distribution and classification loss together in a single loss expression.
He et al., “Deep Mixture Density Network for Probabilistic Object Detection,” 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), October 25-29, 2020, Las Vegas, NV, USA (Virtual) teaches a multi-headed output (see FIG. 2) including classification and MDN outputs, similar to Varamesh.
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to YAO DAVID HUANG whose telephone number is (571)270-1764. The examiner can normally be reached Monday - Friday 9:00 am - 5:30 pm.
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, Miranda Huang can be reached at (571) 270-7092. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/Y.D.H./Examiner, Art Unit 2124
/MIRANDA M HUANG/Supervisory Patent Examiner, Art Unit 2124
Prosecution Insights