DETAILED ACTION
This action is responsive to the application filed on 04/29/2026. Claims 1-4,6-7,9-14,16-17 and 19-20 are pending and have been examined. This action is Final.
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 .
Priority
Applicant’s claim for the benefit of a prior-filed application under 35 U.S.C. 119(e) or under 35 U.S.C.
120, 121, 365(c), or 386(c) is acknowledged.
Response to Arguments
Argument 1: The applicant argues that the amended claims are not directed to an abstract idea, but instead recite a specific technological improvement to machine-learning-based time series prediction. According to the applicant, the claims define a particular predictive architecture that preprocesses historical and current data, trains a model incorporating time-series characteristics and instability information, and predicts both a future state and associated reliability through a structured sequence of algorithmic operations. The applicant further contends that the additional limitations, including model structure creation, multiple prediction points-in-time, environment condition modeling, variation-rate estimation, instability processing, prediction basis generation, and reliability assessment, are not merely mathematical concepts or mental processes, but instead improve the functioning of predictive systems by producing more accurate, reliable, and interpretable forecasting. Accordingly, the applicant asserts that the claims integrate any mathematical concepts into a practical application and recite significantly more than an abstract idea under the Alice/Mayo framework.
Examiner Response to Argument 1: The Examiner has fully considered Applicant’s arguments but does not find them persuasive. Although the amended claims recite a sequence of operations involving preprocessing historical and current data, creating and executing a trained model, processing temporal and environmental features, and calculating prediction instability or reliability, these limitations remain directed to mathematical data analysis and modeling used to generate forecast information. Applicant’s asserted improvements in accuracy, reliability, and interpretability describe desired results of the claimed analysis, but the claims do not recite a particular technical modification to a computer or machine-learning model that produces those results, such as a specifically defined model architecture, data structure, training mechanism, memory-management technique, or computational operation that changes how the computer or model itself functions. Rather, limitations such as “creating a structure of the trained model,” modeling time-series and environment-condition features, and performing a weighted-sum calculation “through deep learning” functionally instruct that the mathematical analysis be carried out using a trained predictive model. When considered individually and as an ordered combination, the additional elements do not impose a meaningful technological limitation on the mathematical concepts or apply them to a particular technological process, but instead use generic machine-learning and computer components to perform the recited calculations in the field of time-series forecasting. Accordingly, the amended claims do not integrate the judicial exception into a practical application under Step 2A, Prong Two, and the additional elements, whether considered separately or in combination, do not amount to significantly more under Step 2B.
Argument 2: The applicant argues that claims 1, 3, 4, 11, 13, 14, and 20 were rejected as anticipated under Section 102 (a)(1) by Zhu, an Uber time series prediction reference, while various dependent claims were rejected as obvious under Section 103 over Zhu combined with Rubanova, Rangapuram, Bishop, and or Lim in different combinations. The applicant traversed all of these rejections on the basis that, even in combination, the cited references fail to disclose or suggest every claimed element as newly amended. Specifically, the applicant argues Zhu does not disclose processing a prediction point in time of the past state data, applying a feature to the past state data by modeling an environment condition feature, or the specific instability calculation now recited. The applicant further distinguishes the claimed instability calculation, a structure in which multiple discrete instability factors such as time series, point in time, and distribution complexity instability are individually derived and then combined via a weighted sum, from Zhu's general concept of predictive uncertainty such as model disagreement or intrinsic noise, and from Lim's weighted combination, which the applicant characterizes as relating to feature combination based on variable importance rather than a weighted sum instability structure. Since independent claims 11 and 20 were amended in parallel with claim 1, the applicant argues the cited art fails to support rejection of the independent claims as amended, and by extension fails to support the dependent claim rejections, requesting withdrawal of all Section 102 and 103 rejections.
Examiner Response to Argument 2: The Examiner has fully considered Applicant’s arguments but does not find them persuasive. The present rejection does not maintain the prior anticipation rejection based solely on Zhu for the newly amended independent claims. Applicant canceled claims 5, 8, 15, and 18 and incorporated their limitations into independent claims 1 and 11. Accordingly, the present obviousness rejection maintains the same allocation of teachings set forth in the non-final Office action mailed December 29, 2025: Zhu teaches the underlying time-series prediction method and apparatus and the calculation of prediction instability generally; Rangapuram teaches creating the trained-model structure, reflecting time-series features in the trained model, processing a prediction point-in-time, and applying environmental or contextual covariates; and Lim teaches weighting processed quantities and combining the weighted quantities. Applicant’s contention that Zhu alone does not disclose the limitations incorporated from former claims 5 and 15 does not address the combined teachings of Zhu and Rangapuram. Further, Applicant’s characterization that the amended claims require multiple discrete instability factors to be separately derived and then combined is not commensurate with the claim language, which recites a weighted sum corresponding to “at least one of” time-series instability, point-in-time instability, and distribution-complexity instability and therefore does not require all three instability types to be separately calculated or combined. Although Lim applies its weighting technique to processed features, the rejection relies on Lim for the known deep-learning technique of assigning weights to corresponding processed quantities and combining those weighted quantities, while Zhu supplies the uncertainty- or instability-related quantities to which the technique is applied. The rejection therefore relies on the references for the same respective teachings relied upon in the prior action, and the amendments do not overcome the combined teachings of Zhu, Rangapuram, and Lim. Claim 11 recites the corresponding limitations in apparatus form, and claim 20 is separately mapped with the same substantive allocation of teachings and its additional transceiver, processor, and memory limitations. Accordingly, Applicant’s arguments do not establish that the amended independent claims, or the claims depending therefrom, are patentable over the cited combinations.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition
of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the
conditions and requirements of this title.
Claims 1-4,6-7,9-14,16-17 and 19-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Regarding claim 1,
Step 1 - Is the claim to a process, machine, manufacture, or composition of matter?
Yes, the claim is directed to a method, which falls under the category of process. The claim satisfies Step 1.
Step 2A - Prong 1 - Does the claim recite an abstract idea, law of nature, or natural phenomenon?
(a) “A method of predicting a future state and reliability based on time series data, comprising: … predicting a future state through execution of an algorithm based on the created trained model, the preprocessed current state data, and the preprocessed past state data” -- The limitation is directed predicting states through executed algorithms based on a trained model and past data. The limitation is directed to a process that can be performed in the human mind using evaluation, observation, and judgement, and thus the limitation is directed to a mental process.
(b) “preprocessing past state data; … preprocessing current state data” -- The limitations are directed to manipulating and conditioning data values according to prescribed rules, such as removing outliers, transforming data values, filling missing values, or determining characteristics of a time-series sequence. Such data conditioning involves mathematical relationships and calculations performed on data, and thus the limitations recite mathematical concepts.
(c) “executing an algorithm reflecting a time series feature of the preprocessed past state data in the trained model; processing a prediction point-in-time of the preprocessed past state data; applying a feature corresponding to the preprocessed past state data by modeling an environment condition feature” -- The limitations are directed to algorithmically identifying, representing, and applying time-series, temporal, and environmental variables within a predictive model. Processing a temporal point, modeling an environmental condition, and applying the resulting feature to data constitute mathematical modeling and analysis of data according to mathematical relationships, and thus recite mathematical concepts.
(d) “calculating instability in the course of predicting the future state based on the time series feature of the preprocessed past state data, wherein the instability is calculated using a weighted sum corresponding to at least one of time series instability, point-in-time instability and distribution complexity instability through deep learning” --The limitation expressly requires calculating an instability value using a weighted sum. A weighted sum involves multiplying one or more input values by corresponding weights and combining the resulting weighted values. The limitation therefore expressly recites a mathematical calculation and relationship and falls within the mathematical-concepts grouping of abstract ideas. The recitation that the calculation is performed “through deep learning” identifies the computational technique used to perform or implement the calculation but does not remove the underlying weighted-sum calculation from the mathematical-concepts grouping.
There are no elements to be evaluated under Step 2A Prong 2 and Step 2B.
Thus, claim 1 is non-patent eligible. Claim 11 has analogous claim limitations to claim 1 aside from the claim type, and thus the rejection above applies to both. Majority of claim 20 is analogous to claim 1 and 11, aside from one limitation, see below:
“a transceiver transmitting and receiving past data and current data to and from an external device;”-- This limitation recites a transceiver transmitting/receiving gathered and current data from an external device. The limitation is an additional elements under Step 2A Prong 2 and Step 2B as an insignificant, extra-solution activity that cannot be integrated to a practical application (see MPEP 2106.05(g)). Furthermore, under Step 2B, the act of transmitting/receiving data over a network/device is a well-understood, routine, and conventional activity (WURC) that cannot provide significantly more than the judicial exception (see MPEP 2106.05(d)(II)). Aside from this limitation, the same rejection can be applied.
Regarding claim 2,
Step 1 - Is the claim to a process, machine, manufacture, or composition of matter?
Yes, the claim is directed to a method, which falls under the category of process. The claim satisfies Step 1.
Step 2A - Prong 1 - Does the claim recite an abstract idea, law of nature, or natural phenomenon?
“The method according to claim 1, wherein the step of preprocessing past state data comprises: removing outliers from past data; and calculating a time series length of the past data.” -- The limitation is directed to the step of the preprocessing of the state data will comprise of removing outliers from data, and then calculating a time series length of the past data. The limitation is directed to the use of a mathematical concept and calculations, and thus the limitation is directed to math.
There are no elements to be evaluated under Step 2A Prong 2 and Step 2B.
Thus, claim 2 is non-patent eligible. Claim 12 has analogous claim limitations to claim 2 aside from the claim type and applying the time series to a processing unit (mere instructions to apply onto a computer). The analogous limitations also are merely just applied to a computer (the device and length calculator) which are additional elements that are not practical applications/significantly more than the judicial exception (see MPEP 2106.05(f)).
Regarding claim 3,
Step 1 - Is the claim to a process, machine, manufacture, or composition of matter?
Yes, the claim is directed to a method, which falls under the category of process. The claim satisfies Step 1.
Step 2A - Prong 1 - Does the claim recite an abstract idea, law of nature, or natural phenomenon?
“creating a trained model by reflecting instability in the created structure of the trained model.” -- The limitation is directed to creating a trained model by reflecting in stability in the trained model structure that was created. Finding instability in a model’s structure is a process that can be performed in the human mind using evaluation, observation, and judgement, and thus the limitation is directed to a mental process.
Step 2A - Prong 2 and Step 2B - Does the claim recite additional elements that integrate the judicial exception into a practical application and amount to significantly more than the abstract idea itself?
“The method according to claim 1, wherein the step of creating a trained model comprises: adjusting a training direction of the trained model based on the preprocessed past state data; creating a structure of the trained model based on at least one of a number of training repetition times, a model size, and an algorithm;” -- The limitation recites the steps of creating a trained model that comprises adjusting a training direction of a model based on past, gathered data, and creating a model based on repetitive training times, a size, and an algorithm. The limitation is directed to an insignificant, extra-solution activity that cannot be integrated to a practical application (see MPEP 2106.05(g)). Furthermore, under Step 2B, the act of adjusting training based of past gathered data and repetitive training is a well-understood, routine, and conventional activity (WURC), and thus it cannot provide significantly more than the judicial exception (see MPEP 2106.05(d)(II)).
Thus, claim 3 is non-patent eligible. Claim 13 has analogous claim limitations to claim 3 aside from the claim type, and thus the rejection above applies to both.
Regarding claim 4,
Step 1 - Is the claim to a process, machine, manufacture, or composition of matter?
Yes, the claim is directed to a method, which falls under the category of process. The claim satisfies Step 1.
Step 2A - Prong 1 - Does the claim recite an abstract idea, law of nature, or natural phenomenon?
“the trained model is created using at least one of percentage instability and instability with reference to a specific critical point.” --The limitation is directed to the trained model being created using a percentage of instability and reference to a specific critical point, which is calculated. The limitation is directed to a use of a mathematical relationship/calculation, and thus the limitation is directed to math.
Step 2A - Prong 2 and Step 2B - Does the claim recite additional elements that integrate the judicial exception into a practical application and amount to significantly more than the abstract idea itself?
“wherein, in generation of the trained model by reflecting instability in the created structure of the trained model,” --The limitation recites simple further limits to how the instability/reflected instability is represented in the trained model, and it does not integrate to a practical application, nor does it provide significantly more than the judicial exception (see MPEP 2106.05(h)).
Thus, claim 4 is non-patent eligible. Claim 14 is analogous to claim 4, aside from claim type, and thus the same rejection above can be applied to both claims.
Regarding claim 6,
Step 1 - Is the claim to a process, machine, manufacture, or composition of matter?
Yes, the claim is directed to a method, which falls under the category of process. The claim satisfies Step 1.
Step 2A - Prong 1 - Does the claim recite an abstract idea, law of nature, or natural phenomenon?
“creating a complexity distribution of the preprocessed past state data; creating a complexity distribution sample based on the created complexity distribution; and creating future state data based on the created complexity distribution sampling.” -- The limitation recites creating complexity distribution samples based on past distribution, then same for future state data. The limitation is directed to the use a mathematical calculation/operations, and thus the limitation is directed to math.
Step 2A - Prong 2 and Step 2B - Does the claim recite additional elements that integrate the judicial exception into a practical application and amount to significantly more than the abstract idea itself?
“The method according to claim 1, further comprising: applying a suitable feature to the preprocessed past state data by modeling the environment condition feature;” -- The limitation recited the method will further comprise applying a suitable feature to the preprocessed past state data by modeling the feature, and it is merely limiting the field of use of the claim, which is not integrating to a practical application, nor provides significantly more than the judicial exception (see MPEP 2106.05(h)).
Thus, claim 6 is non-patent eligible. Claim 16 is analogous to claim 6, aside from claim type, and thus the same rejection above can be applied to both claims.
Regarding claim 7,
Step 1 - Is the claim to a process, machine, manufacture, or composition of matter?
Yes, the claim is directed to a method, which falls under the category of process. The claim satisfies Step 1.
Step 2A - Prong 1 - Does the claim recite an abstract idea, law of nature, or natural phenomenon?
“The method according to claim 1, wherein the step of processing a prediction point-in-time of the preprocessed past state data comprises: training a function for estimation of a variation rate of input data through deep learning; and calculating a variation estimation function depending on the prediction point-in-time using the function.” -- The limitation is directed to calculating a variation est. function depending on a predicted point-in time using a function. The limitation is directed to the use of mathematical concept/operation/calculation, as well as recites a process that can be performed in the human mind using evaluation, observation, and judgement, with aid of paper, thus the limitation can be directed to both mental process AND math.
There are no elements to be evaluated under Step 2A Prong 2 and Step 2B.
Thus, claim 7 is non-patent eligible. Claim 17 is analogous to claim 7, aside from claim type, and thus the same rejection above can be applied to both claims.
Regarding claim 9,
Step 1 - Is the claim to a process, machine, manufacture, or composition of matter?
Yes, the claim is directed to a method, which falls under the category of process. The claim satisfies Step 1.
Step 2A - Prong 1 - Does the claim recite an abstract idea, law of nature, or natural phenomenon?
“ The method according to claim 1, wherein the step of predicting a future state comprises: calculating a prediction point-in-time by receiving a future point-in-time that a user wants to know; and predicting a future state of the user through execution of an algorithm based on the prediction point-in-time and the trained model.” -- The limitation is directed to steps to predict a future state b either calculating a prediction point-in-time once a user-requested point-in-time is received in the future/through the execution of an algorithm. The limitation is directed to a process that can be performed in the human mind using evaluation, observation, and judgment, with aid of paper, thus the limitation can be directed to mental process.
There are no elements to be evaluated under Step 2A Prong 2 and Step 2B.
Thus, claim 9 is non-patent eligible. Claim 19 is analogous to claim 9, aside from claim type, and thus the same rejection above can be applied to both claims.
Regarding claim 10,
Step 1 - Is the claim to a process, machine, manufacture, or composition of matter?
Yes, the claim is directed to a method, which falls under the category of process. The claim satisfies Step 1.
There are no elements to be evaluated under Step 2A Prong 1.
Step 2A - Prong 2 and Step 2B - Does the claim recite additional elements that integrate the judicial exception into a practical application and amount to significantly more than the abstract idea itself?
“The method according to claim 9, wherein the step of predicting the future state comprises: calculating at least one of reliability of the future state, a prediction basis of the future state, and instability.” -- The limitation recites predicting the future state will comprise now reliability, a prediction basis of the future state, and instability. The limitation merely limits to a field of use/environment, and thus it does not integrate to a practical application, nor provides significantly more than the judicial exception (see MPEP 2106.05(h)).
Thus, claim 10 is non-patent eligible.
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 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.
Claim(s) 1, 3-4, 7, 11, 13-14, 17, and 20 are rejected under 35 U.S.C. 103 as being unpatentable over NPL reference “Deep and Confident Prediction for Time Series at Uber”, by Zhu et. al. (referred herein as Zhu) in view of NPL reference “Deep state space models for time series forecasting.”, by Rangapuram et. al. (referred herein as Rangapuram) further in view of NPL reference “Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting”, by Lim et al. (referred herein as Lim).
Regarding claim 1, Zhu teaches:
A method of predicting a future state and reliability based on time series data, ([Zhu, page 1, Abstract] “we propose a novel end-to-end Bayesian deep model that provides time series prediction along with uncertainty estimation,” wherein the examiner interprets “time series prediction along with uncertainty estimation” to be the same as predicting a future state and reliability based on time series data because both are directed to forecasting future values from sequential data while determining the uncertainty or reliability associated with the forecast.)
comprising: preprocessing past state data; ([Zhu, page 5, sec. 4.1.1] “The raw data are log-transformed to alleviate exponential effects. Next, within each sliding window, the first day is subtracted from all values, so that trends are removed and the neural network is trained for the incremental value,” wherein the examiner interprets log-transforming the raw data and removing trends from the sliding-window data to be the same as preprocessing past state data because both are directed to transforming historical time-series data before model training and prediction.)
creating a trained model through execution of an algorithm based on the preprocessed past state data; ([Zhu, page 4, sec. 3.2.1] “Prior to fitting the prediction model, we first conduct a pre-training step to fit an encoder that can extract useful and representative embeddings from a time series” AND [Zhu, page 5, sec. 3.2.2] “a prediction network is trained to forecast the next one or more timestamps using the learned embedding as features,” wherein the examiner interprets pre-training an encoder and training a prediction network using embeddings extracted from historical time-series data to be the same as creating a trained model through execution of an algorithm based on the preprocessed past state data because both are directed to algorithmically training a model using processed historical observations.)
preprocessing current state data; and ([Zhu, page 3, sec. 3] “Given a trained neural network … as well as a new sample x*, our goal is to evaluate the uncertainty of the model prediction” AND [Zhu, page 5, sec. 4.1.1] “The raw data are log-transformed” and “At test time, it is straightforward to revert these transformations to obtain predictions at the original scale,” wherein the examiner interprets receiving and processing a new test-time sample using the disclosed data transformations to be the same as preprocessing current state data because both are directed to preparing newly received time-series observations for model inference.)
predicting a future state through execution of an algorithm based on the created trained model, the preprocessed current state data, and the preprocessed past state data, ([Zhu, page 5, secs. 3.2.2, 3.2.3 and 4.1.1] “a prediction network is trained to forecast the next one or more timestamps using the learned embedding as features,” AND “After the full model is trained, the inference stage involves only the encoder and the prediction network,” AND “each sliding window contains the previous 28 days as input, and aims to forecast the upcoming day,” wherein the examiner interprets using the trained encoder and prediction network with a sliding window of historical and test-time data to forecast an upcoming timestamp to be the same as predicting a future state through execution of an algorithm based on the trained model, preprocessed current state data, and preprocessed past state data.)
calculating instability in the course of predicting the future state based on the time series feature of the preprocessed past state data. ([Zhu, page 3, sec. 3] “our goal is to evaluate the uncertainty of the model prediction” AND [Zhu, page 3, sec. 3.1] “the variance of the prediction distribution quantifies the prediction uncertainty, which can be further decomposed using law of total variance” AND [Zhu, page 3, sec. 3.1.1] “the model uncertainty can be approximated by the sample variance,” wherein the examiner interprets evaluating and quantifying prediction uncertainty by variance and sample variance to be the same as calculating instability in the course of predicting the future state based on the time-series feature of the preprocessed past state data because both compute a quantitative measure of variability or uncertainty associated with a prediction generated from temporal input data.)
an instability corresponding to at least one of time series instability, point-in-time instability, and distribution complexity instability ([Zhu, page 1, sec. 1] “The third uncertainty from model misspecification … captures the scenario where the testing samples come from a different population than the training set” AND [Zhu, page 3, sec. 3.1.2] “we would like to capture the uncertainty when predicting unseen samples with very different patterns from the training data set,” wherein the examiner interprets model-misspecification uncertainty to be the same as distribution-complexity instability because both concern prediction instability caused by differences between the distribution or patterns of prediction data and the data used to train the model.)
through deep learning. ([Zhu, page 4, sec. 3.1.2] “During this feedforward pass, MC dropout is applied to all layers in both the encoder g and the prediction network h” AND “variational dropout for recurrent neural networks … is applied to the LSTM layers in the encoder, and regular dropout … is applied to the prediction network,” wherein the examiner interprets applying Monte Carlo dropout to the LSTM encoder and neural-network prediction network to calculate model and misspecification uncertainty to be the same as calculating instability through deep learning because both use a trained multilayer neural-network architecture to calculate prediction uncertainty or instability.)
Zhu does not teach wherein the step of predicting the future state through execution of the algorithm includes: creating a structure of the trained model; executing an algorithm reflecting a time series feature of the preprocessed past state data in the trained model; processing a prediction point-in-time of the preprocessed past state data; and applying a feature corresponding to the preprocessed past state data by modeling an environment condition feature.
Rangapuram teaches:
wherein the step of predicting the future state through execution of the algorithm includes: creating a structure of the trained model; ([Rangapuram, page 4, sec. 4] “We parameterize the mapping Ψ from covariates to state space model parameters using a deep recurrent neural network (RNN)” AND “a multi-layer recurrent neural network with LSTM cells and parameters Φ computes a representation of the features via a recurrent function h,” wherein the examiner interprets parameterizing the mapping from covariates to state-space-model parameters using a multilayer recurrent neural network with LSTM cells to be the same as creating a structure of the trained model because both define the neural-network architecture and model components used for prediction.)
executing an algorithm reflecting a time series feature of the preprocessed past state data in the trained model; ([Rangapuram, page 3, sec. 3] “SSMs model the temporal structure of the data via a latent state l_t ∈ R^L that can be used to encode time series components such as level, trend, and seasonality patterns” AND [Rangapuram, page 4, sec. 4] “The real-valued output vector of the last LSTM layer is then mapped to the parameters Θ_t^(i) of the state space model,” wherein the examiner interprets encoding level, trend, and seasonality patterns in a latent state and mapping the LSTM output to state-space-model parameters to be the same as executing an algorithm reflecting a time-series feature of the preprocessed past state data in the trained model because both incorporate temporal characteristics of historical data into the trained prediction model.)
processing a prediction point-in-time of the preprocessed past state data; ([Rangapuram, page 3, sec. 3] “The time point T_i + 1 is referred to as forecast start time and τ ∈ N>0 is the forecast horizon” AND [Rangapuram, pages 5-6, sec. 4.2] “we unroll the RNN for the prediction range t = T_i + 1, …, T_i + τ … then generate the prediction samples by recursively applying above equations,” wherein the examiner interprets defining a forecast-start time and forecast horizon and unrolling the RNN over the corresponding prediction range to be the same as processing a prediction point-in-time of the preprocessed past state data because both process defined future temporal positions based on historical time-series observations.)
applying a feature corresponding to the preprocessed past state data by modeling an environment condition feature; and ([Rangapuram, page 2, sec. 3] “let {x^(i)_(1:T_i+τ)}_(i=1)^N be a set of associated, time-varying covariate vectors” AND [Rangapuram, page 4, sec. 4.1] “The covariates (features) can be time dependent (e.g. product price or a set of dummy variables indicating day-of-week) or time independent (e.g., product brand, category etc.),” wherein the examiner interprets associated time-varying or time-independent covariates representing price, day-of-week, brand, or category to be the same as applying a feature corresponding to the preprocessed past state data by modeling an environment condition feature because both incorporate contextual or environmental conditions affecting the time series.)
Zhu and Rangapuram do not teach wherein the instability is calculated using a weighted sum.
Lim teaches using a weighted sum. ([Lim, page 1752, sec. 4.2] “Processed features are then weighted by their variable selection weights and combined,” wherein the examiner interprets weighting processed quantities by corresponding variable-selection weights and combining the weighted quantities to be the same as using a weighted sum because both multiply corresponding quantities by weights and aggregate the weighted quantities into a combined result.)
Zhu, Rangapuram, Lim, and the instant application are analogous art because they are all directed to deep-learning-based time-series prediction and the quantitative processing of features or uncertainty-related information used to generate or evaluate future-state predictions.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the neural-network time-series prediction method disclosed by Zhu to include the deep recurrent state-space-model structure, temporal-feature encoding, forecast-time processing, and associated time-varying covariates disclosed by Rangapuram. One would be motivated to do so to effectively model temporal structure and contextual information over a defined forecast horizon and thereby improve the quality and robustness of future-state predictions, as suggested by Rangapuram ([Rangapuram, page 3, sec. 3] “SSMs model the temporal structure of the data via a latent state” AND [Rangapuram, page 2, sec. 3] “Our goal is to produce a set of probabilistic forecasts”).
It would have further been obvious to a person of ordinary skill in the art before the effective filing date of the invention to include the “weighted by their variable selection weights and combined” technique disclosed by Lim. One would be motivated to do so to efficiently aggregate multiple uncertainty-related quantities into a single instability measure while reducing the influence of less relevant information, as suggested by Lim ([Lim, page 1752, sec. 4.2] “variable selection also allows TFT to remove any unnecessary noisy inputs which could negatively impact performance”).Claim 11 is analogous to claim 1, aside from claim type and minute differences, and thus faces the same rejection as above. Claim 20 is partially analogous but differs enough to warrant its own rejection.
Regarding claim 3, Zhu, Rangapuram, and Lim teaches The method according to claim 1, (see rejection of claim 1).
Zhu further teaches:
the step of creating a trained model comprises: adjusting a training direction of the trained model based on the preprocessed past state data; ([Zhu, page 3, sec 3.1] “Given a set of N observations X = {x1, ..., xN } and Y = {y1, ..., yN }, Bayesian inference aims at finding the posterior distribution over model parameters p(W | X, Y )”, wherein the examiner interprets “Bayesian inference aims at finding the posterior distribution over model parameters” to be the same as adjusting a training direction of the trained model based on the preprocessed past state data because they are both directed to updating and optimizing model parameters based on the training observations.)
creating a structure of the trained model based on at least one of a number of training repetition times, a model size, and an algorithm; and ([Zhu, page 5, sec 4.1.1] “The encoder-decoder is constructed with two-layer LSTM cells, with 128 and 32 hidden states, respectively. The prediction network has three fully connected layers with tanh activation, with 128, 64, and 16 hidden units, respectively”, wherein the examiner interprets “two-layer LSTM cells, with 128 and 32 hidden states” and “three fully connected layers with tanh activation, with 128, 64, and 16 hidden units” to be the same as creating a structure of the trained model based on at least one of the number of training repetition times, a model size, and an algorithm because they are both directed to defining the architecture and size parameters of the neural network model.)
creating a trained model by reflecting instability in the created structure of the trained model. ([Zhu, page 1, sec 1] “Under this framework, the prediction uncertainty can be decomposed into three types: model uncertainty, inherent noise, and model misspecification” AND [Zhu, page 2, sec 2.2] “stochastic dropouts are applied after each hidden layer, and the model output can be approximately viewed as a random sample generated from the posterior predictive distribution”, wherein the examiner interprets “prediction uncertainty can be decomposed into three types: model uncertainty, inherent noise, and model misspecification” and “stochastic dropouts are applied after each hidden layer” to be the same as creating a trained model by reflecting instability in the created structure of the trained model because they are both directed to incorporating uncertainty quantification mechanisms into the model architecture to capture prediction variability.)
Regarding claim 4, Zhu, Rangapuram, and Lim teaches The method according to claim 1, (see rejection of claim 1).
Zhu further teaches:
wherein, in generation of the trained model by reflecting instability in the created structure of the trained model, ([Zhu, page 4, sec 3.1.2] “Here, we take a principled approach by connecting the encoder, g(·), with a prediction network, h(·), and treat them as one large network f = h(g(·)) during inference” AND [Zhu, page 4, sec 3.1.2] “During this feedforward pass, MC dropout is applied to all layers in both the encoder g and the prediction network h”, wherein the examiner interprets “treat them as one large network” and “MC dropout is applied to all layers in both the encoder g and the prediction network h” to be the same as in generation of the trained model by reflecting instability in the created structure of the trained model because they are both directed to building a model architecture that inherently incorporates uncertainty estimation mechanisms within its structural design.)
the trained model is created using at least one of percentage instability ([Zhu, page 3, sec 3] “Specifically, we would like to quantify the prediction standard error, η, so that an approximate α-level prediction interval can be constructed by [ŷ* - zα/2η, ŷ* + zα/2η]” AND [Zhu, page 6, sec 4.1.3] “Table 2 reports the empirical coverage of the 95% predictive intervals”, wherein the examiner interprets “α-level prediction interval” and “empirical coverage of the 95% predictive intervals” to be the same as percentage instability because they are both directed to expressing uncertainty as a percentage-based confidence level that quantifies the range within which predictions are expected to fall.)
and instability with reference to a specific critical point. ([Zhu, page 6, sec 4.1.3] “One important use-case of the uncertainty estimation is to provide insight for unusual patterns in the time series” AND [Zhu, page 7, sec 4.2] “A natural approach is to trigger an alarm when the observed value falls outside of the 95% predictive interval”, wherein the examiner interprets “trigger an alarm when the observed value falls outside of the 95% predictive interval” to be the same as instability with reference to a specific critical point because they are both directed to identifying when values exceed a defined threshold or boundary that serves as a reference point for determining abnormal or unstable conditions.)
Regarding claim 7, Zhu, Rangapuram, and Lim teaches The method according to claim 1, (see rejection of claim 1).
Rangapuram further teaches:
wherein the step of processing a prediction point-in-time of the preprocessed past state data comprises: training a function for estimation of a variation rate of input data through deep learning; ([Rangapuram, page 3, sec 3] “the latent state lt-1 maintains information about level, trend, and seasonality patterns and evolves by way of a deterministic transition matrix Ft and a random innovation gtεt” , [Rangapuram, page 3, sec 3] “The structure of the transition matrix Ft and innovation strength gt determine which kind of time series patterns are encoded by the latent state lt” AND [Rangapuram, page 4, sec 4.1] “The model parameters Φ are learned by maximizing the probability of observing the data”, wherein the examiner interprets “evolves by way of a deterministic transition matrix Ft and a random innovation gtεt” and “transition matrix Ft and innovation strength gt determine which kind of time series patterns” and “model parameters Φ are learned by maximizing the probability” to be the same as training a function for estimation of a variation rate of input data through deep learning because they are both directed to learning a function through neural network optimization that captures how the state changes or varies over time.)
and calculating a variation estimation function depending on the prediction point-in-time using the function. ([Rangapuram, page 5, sec 4.2] “starting with sample T ∼ p(T |z1:T ), we recursively apply yT+t = a>T+tT+t-1 + bT+t, t = 1, . . . τ…lT+t ∼ FT+tT+t-1 + gT+tεT+t, εT+t ∼ N (0, 1), t = 1, . . . τ - 1” AND [Rangapuram, page 6, sec 4.2] “we unroll the RNN for the prediction range t = Ti + 1, . . . , Ti + τ and obtain Θ(i)Ti+1:Ti+τ, then generate the prediction samples by recursively applying above equations”, wherein the examiner interprets “recursively apply yT+t = a>T+tT+t-1 + bT+t” and “lT+t ∼ FT+tT+t-1 + gT+tεT+t” and “unroll the RNN for the prediction range” to be the same as calculating a variation estimation function depending on the prediction point-in-time using the function because they are both directed to applying the learned transition dynamics to compute state evolution and predictions at each specified future time point in the forecast horizon.)
Zhu, Rangapuram, Lim, and the instant application are analogous art because they are all directed to predicting future states of time-series data using learned models that account for temporal variation and uncertainty across a forecast horizon.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the method claim 1 disclosed by Zhu, Rangapuram, and Lim to include the use of a transition matrix disclosed by Rangapuram. One would be motivated to do so to more accurately and effectively model temporal variation across future prediction points, as suggested by Rangapuram ([Rangapuram, page 3, sec 3] “the latent state lt-1maintains information about level, trend, and seasonality patterns and evolves by way of a deterministic transition matrix F_t and a random innovation 𝑔𝑡𝜀𝑡”).
Regarding claim 13, Zhu, Rangapuram, and Lim teaches The apparatus according to claim 11, (see rejection of claim 1).
Zhu further teaches:
wherein the future state prediction model-training unit comprises: ([Zhu, page 4, sec 3.2] “The complete architecture of the neural network is shown in Figure 1. The network contains two major components: (i) an encoder-decoder framework that captures the inherent pattern in the time series, which is learned during pre-training step, and (ii) a prediction network”, wherein the examiner interprets “complete architecture of the neural network” and “network contains two major components” to be the same as the future state prediction model-training unit comprises because they are both directed to a system comprising multiple functional components for training a prediction model.)
a multiple points-in-time generator adjusting a training direction of the trained model based on the preprocessed past state data; ([Zhu, page 5, sec 4.1.1 ] “Samples are constructed using a sliding window with step size one, where each sliding window contains the previous 28 days as input” AND [Zhu, page 3, sec 3.1] “Given a set of N observations X = {x1, ..., xN } and Y = {y1, ..., yN }, Bayesian inference aims at finding the posterior distribution over model parameters p(W | X, Y)”, wherein the examiner interprets “Samples are constructed using a sliding window” to be the same as a multiple points-in-time generator because they are both directed to generating multiple temporal samples from the time series, and “Bayesian inference aims at finding the posterior distribution over model parameters” to be the same as adjusting a training direction of the trained model based on the preprocessed past state data because they are both directed to updating and optimizing model parameters based on the training observations.)
a time series model-training unit creating a structure of the trained model based on at least one of the number of training repetition times, a model size, and an algorithm; and ([Zhu, page 5, sec 4.1.1] “The encoder-decoder is constructed with two-layer LSTM cells, with 128 and 32 hidden states, respectively. The prediction network has three fully connected layers with tanh activation, with 128, 64, and 16 hidden units, respectively”, wherein the examiner interprets “two-layer LSTM cells, with 128 and 32 hidden states” and “three fully connected layers with tanh activation, with 128, 64, and 16 hidden units” to be the same as a time series model-training unit creating a structure of the trained model based on at least one of the number of training repetition times, a model size, and an algorithm because they are both directed to defining the architecture and size parameters of the neural network model.)
an instability calculator creating a trained model by reflecting instability in the created structure of the trained model. ([Zhu, page 1, sec 1] “Under this framework, the prediction uncertainty can be decomposed into three types: model uncertainty, inherent noise, and model misspecification” AND [Zhu, page 2, sec 2.2] “stochastic dropouts are applied after each hidden layer, and the model output can be approximately viewed as a random sample generated from the posterior predictive distribution”, wherein the examiner interprets “prediction uncertainty can be decomposed into three types: model uncertainty, inherent noise, and model misspecification” and “stochastic dropouts are applied after each hidden layer” to be the same as an instability calculator creating a trained model by reflecting instability in the created structure of the trained model because they are both directed to incorporating uncertainty quantification mechanisms into the model architecture to capture prediction variability.)
Regarding claim 14, Zhu, Rangapuram, and Lim teaches The apparatus according to claim 13, (see rejection of claim 13).
Zhu further teaches:
wherein the instability calculator creates the trained model ([Zhu, page 4, sec 3.1.2] “Here, we take a principled approach by connecting the encoder, g(·), with a prediction network, h(·), and treat them as one large network f = h(g(·)) during inference” AND [Zhu, page 4, sec 3.1.2] “During this feedforward pass, MC dropout is applied to all layers in both the encoder g and the prediction network h”, wherein the examiner interprets “treat them as one large network” and “MC dropout is applied to all layers in both the encoder g and the prediction network h” to be the same as the instability calculator creates the trained model because they are both directed to building a model architecture that inherently incorporates uncertainty estimation mechanisms within its structural design.)
using at least one of percentage instability ([Zhu, page 3, sec 3] “Specifically, we would like to quantify the prediction standard error, η, so that an approximate α-level prediction interval can be constructed by [ŷ* - zα/2η, ŷ* + zα/2η]” AND [Zhu, page 6, sec 4.1.3] “Table 2 reports the empirical coverage of the 95% predictive intervals”, wherein the examiner interprets “α-level prediction interval” and “empirical coverage of the 95% predictive intervals” to be the same as percentage instability because they are both directed to expressing uncertainty as a percentage-based confidence level that quantifies the range within which predictions are expected to fall.)
and instability with reference to a specific critical point. ([Zhu, page 6, sec 4.1.3] “One important use-case of the uncertainty estimation is to provide insight for unusual patterns in the time series” AND [Zhu, page 7, sec 4.2] “A natural approach is to trigger an alarm when the observed value falls outside of the 95% predictive interval”, wherein the examiner interprets “trigger an alarm when the observed value falls outside of the 95% predictive interval” to be the same as instability with reference to a specific critical point because they are both directed to identifying when values exceed a defined threshold or boundary that serves as a reference point for determining abnormal or unstable conditions.)
Regarding claim 17, Zhu, Rangapuram, and Lim teach The apparatus according to claim 11, (see rejection of claim 11).
Rangapuram further teaches:
wherein the point-in-time feature processing device trains a function for estimation of a variation rate of input data through deep learning ([Rangapuram, page 3, sec 3] “the latent state lt-1 maintains information about level, trend, and seasonality patterns and evolves by way of a deterministic transition matrix Ft and a random innovation gtεt” AND [Rangapuram, page 3, sec 3] “The structure of the transition matrix Ft and innovation strength gt determine which kind of time series patterns are encoded by the latent state lt” AND [Rangapuram, page 4, sec 4.1] “The model parameters Φ are learned by maximizing the probability of observing the data”, wherein the examiner interprets “evolves by way of a deterministic transition matrix Ft and a random innovation gtεt” and “transition matrix Ft and innovation strength gt determine which kind of time series patterns” and “model parameters Φ are learned by maximizing the probability” to be the same as the point-in-time feature processing device trains a function for estimation of a variation rate of input data through deep learning because they are both directed to a processing component that learns a function through neural network optimization that captures how the state changes or varies over time.)
and calculates a variation estimation function depending on a prediction point-in-time using the trained function for estimation of a variation rate of input data. ([Rangapuram, page 5, sec 4.2] “starting with sample T ∼ p(T |z1:T ), we recursively apply yT+t = a>T+tT+t-1 + bT+t, t = 1, . . . τ…lT+t ∼ FT+tT+t-1 + gT+tεT+t, εT+t ∼ N (0, 1), t = 1, . . . τ - 1” AND [Rangapuram, page 6, sec 4.2] “we unroll the RNN for the prediction range t = Ti + 1, . . . , Ti + τ and obtain Θ(i)Ti+1:Ti+τ, then generate the prediction samples by recursively applying above equations”, wherein the examiner interprets “recursively apply yT+t = a>T+tT+t-1 + bT+t” and “lT+t ∼ FT+tT+t-1 + gT+tεT+t” and “unroll the RNN for the prediction range” to be the same as calculates a variation estimation function depending on a prediction point-in-time using the trained function for estimation of a variation rate of input data because they are both directed to a processing component that applies the learned transition dynamics to compute state evolution and predictions at each specified future time point in the forecast horizon.)
Zhu, Rangapuram, Lim, and the instant application are analogous art because they are all directed to predicting future states of time-series data using learned models that account for temporal variation and uncertainty across a forecast horizon.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the method claim 11 disclosed by Zhu, Rangapuram, and Lim to include the use of a transition matrix disclosed by Rangapuram. One would be motivated to do so to more accurately and effectively model temporal variation across future prediction points, as suggested by Rangapuram ([Rangapuram, page 3, sec 3] “the latent state 1t-1 maintains information about level, trend, and seasonality patterns and evolves by way of a deterministic transition matrix F_t and a random innovation 𝑔𝑡𝜀𝑡”).
Regarding claim 20, Zhu teaches:
An apparatus for predicting a future state and reliability based on time series data, ([Zhu, page 1, Abstract] “we propose a novel end-to-end Bayesian deep model that provides time series prediction along with uncertainty estimation,” wherein the examiner interprets “time series prediction along with uncertainty estimation” to be the same as an apparatus for predicting a future state and reliability based on time series data because both are directed to a computer-implemented system that forecasts future values from sequential data while determining the uncertainty or reliability associated with the forecast.)
comprising: a transceiver transmitting and receiving past data and current data to and from an external device; ([Zhu, page 7, sec. 4.2] “At Uber, we track millions of metrics each day to monitor the status of various services across the company” AND “provide real-time anomaly detection and deploy alerts for potential outages and unusual behaviors,” wherein the examiner interprets tracking past and current metrics generated by external services and deploying alerts based on the processed metrics to be the same as a transceiver transmitting and receiving past data and current data to and from an external device because both are directed to receiving time-series data from external systems and transmitting information generated from processing the received time-series data.)
a processor preprocessing at least one of the past state data and the current state data, ([Zhu, page 5, sec. 4.1.1] “The raw data are log-transformed to alleviate exponential effects. Next, within each sliding window, the first day is subtracted from all values, so that trends are removed and the neural network is trained for the incremental value,” wherein the examiner interprets log-transforming the raw data and removing trends from the sliding-window data to be the same as a processor preprocessing at least one of the past state data and the current state data because both are directed to computationally transforming time-series data before model training and prediction.)
creating a trained model through execution of an algorithm based on the preprocessed past state data, ([Zhu, page 4, sec. 3.2.1] “Prior to fitting the prediction model, we first conduct a pre-training step to fit an encoder that can extract useful and representative embeddings from a time series” AND [Zhu, page 5, sec. 3.2.2] “a prediction network is trained to forecast the next one or more timestamps using the learned embedding as features,” wherein the examiner interprets pre-training an encoder and training a prediction network using embeddings extracted from historical time-series data to be the same as a processor creating a trained model through execution of an algorithm based on the preprocessed past state data because both are directed to algorithmically training a model using processed historical observations.)
and predicting a future state through execution of an algorithm based on the created trained model, the preprocessed current state data, and the preprocessed past state data; ([Zhu, page 5, secs. 3.2.2, 3.2.3 and 4.1.1] “a prediction network is trained to forecast the next one or more timestamps using the learned embedding as features,” AND “After the full model is trained, the inference stage involves only the encoder and the prediction network,” AND “each sliding window contains the previous 28 days as input, and aims to forecast the upcoming day,” wherein the examiner interprets using the trained encoder and prediction network with a sliding window of historical and test-time data to forecast an upcoming timestamp to be the same as a processor predicting a future state through execution of an algorithm based on the trained model, preprocessed current state data, and preprocessed past state data.)
and a memory storing the trained model and the future state, ([Zhu, page 3, sec. 3] “Given a trained neural network (f_{\hat W}(\cdot)) where (\hat W) represents the fitted parameters” AND [Zhu, page 4, Algorithm 2] returning a prediction (\hat y^*) and predictive uncertainty (\eta) AND [Zhu, page 7, sec. 4.2.1] “Our model inference is implemented in Go. Our implementation involves efficient matrix manipulation operations,” wherein the examiner interprets retaining the fitted parameters of the trained neural network and the resulting future-state prediction during computer-implemented inference to be the same as a memory storing the trained model and the future state because both are directed to computationally storing trained-model information and a future-state prediction for execution, retrieval, or subsequent use.)
wherein the processor is configured to calculate instability in the course of predicting the future state based on the time series feature of the preprocessed past state data, ([Zhu, page 3, sec. 3] “our goal is to evaluate the uncertainty of the model prediction” AND [Zhu, page 3, sec. 3.1] “the variance of the prediction distribution quantifies the prediction uncertainty, which can be further decomposed using law of total variance” AND [Zhu, page 3, sec. 3.1.1] “the model uncertainty can be approximated by the sample variance,” wherein the examiner interprets a processor evaluating and quantifying prediction uncertainty by variance and sample variance to be the same as a processor configured to calculate instability during future-state prediction based on time-series features because both compute a quantitative measure of variability or uncertainty associated with a prediction generated from temporal input data.)
corresponding to at least one of time series instability, point-in-time instability, and distribution complexity instability ([Zhu, page 1, sec. 1] “The third uncertainty from model misspecification … captures the scenario where the testing samples come from a different population than the training set” AND [Zhu, page 3, sec. 3.1.2] “we would like to capture the uncertainty when predicting unseen samples with very different patterns from the training data set,” wherein the examiner interprets model-misspecification uncertainty to be the same as distribution-complexity instability because both concern prediction instability caused by differences between the distribution or patterns of prediction data and the data used to train the model.)
through deep learning. ([Zhu, page 4, sec. 3.1.2] “During this feedforward pass, MC dropout is applied to all layers in both the encoder g and the prediction network h” AND “variational dropout for recurrent neural networks … is applied to the LSTM layers in the encoder, and regular dropout … is applied to the prediction network,” wherein the examiner interprets applying Monte Carlo dropout to the LSTM encoder and neural-network prediction network to calculate model and misspecification uncertainty to be the same as configuring the processor to calculate instability through deep learning because both use a trained multilayer neural-network architecture to calculate prediction uncertainty or instability.)
Zhu does not teach wherein the processor is configured to create a structure of the trained model; execute an algorithm reflecting a time series feature of the preprocessed past state data in the trained model; process a prediction point-in-time of the preprocessed past state data; and apply a feature corresponding to the preprocessed past state data by modeling an environment condition feature.
Rangapuram teaches:
wherein the processor is configured to create a structure of the trained model; ([Rangapuram, page 4, sec. 4] “We parameterize the mapping Ψ from covariates to state space model parameters using a deep recurrent neural network (RNN)” AND “a multi-layer recurrent neural network with LSTM cells and parameters Φ computes a representation of the features via a recurrent function h,” wherein the examiner interprets parameterizing the mapping from covariates to state-space-model parameters using a multilayer recurrent neural network with LSTM cells to be the same as configuring the processor to create a structure of the trained model because both define the neural-network architecture and model components used for prediction.)
execute an algorithm reflecting a time series feature of the preprocessed past state data in the trained model; ([Rangapuram, page 3, sec. 3] “SSMs model the temporal structure of the data via a latent state l_t ∈ R^L that can be used to encode time series components such as level, trend, and seasonality patterns” AND [Rangapuram, page 4, sec. 4] “The real-valued output vector of the last LSTM layer is then mapped to the parameters Θ_t^(i) of the state space model,” wherein the examiner interprets encoding level, trend, and seasonality patterns in a latent state and mapping the LSTM output to state-space-model parameters to be the same as configuring the processor to execute an algorithm reflecting a time-series feature of the preprocessed past state data in the trained model because both incorporate temporal characteristics of historical data into the trained prediction model.)
process a prediction point-in-time of the preprocessed past state data; ([Rangapuram, page 3, sec. 3] “The time point T_i + 1 is referred to as forecast start time and τ ∈ N>0 is the forecast horizon” AND [Rangapuram, pages 5-6, sec. 4.2] “we unroll the RNN for the prediction range t = T_i + 1, …, T_i + τ … then generate the prediction samples by recursively applying above equations,” wherein the examiner interprets defining a forecast-start time and forecast horizon and unrolling the RNN over the corresponding prediction range to be the same as configuring the processor to process a prediction point-in-time of the preprocessed past state data because both process defined future temporal positions based on historical time-series observations.)
apply a feature corresponding to the preprocessed past state data by modeling an environment condition feature; and ([Rangapuram, page 2, sec. 3] “let {x^(i)_(1:T_i+τ)}_(i=1)^N be a set of associated, time-varying covariate vectors” AND [Rangapuram, page 4, sec. 4.1] “The covariates (features) can be time dependent (e.g. product price or a set of dummy variables indicating day-of-week) or time independent (e.g., product brand, category etc.),” wherein the examiner interprets associated time-varying or time-independent covariates representing price, day-of-week, brand, or category to be the same as configuring the processor to apply a feature corresponding to the preprocessed past state data by modeling an environment condition feature because both incorporate contextual or environmental conditions affecting the time series.)
Zhu and Rangapuram do not teach wherein the instability is calculated using a weighted sum.
Lim teaches using a weighted sum. ([Lim, page 1752, sec. 4.2] “Processed features are then weighted by their variable selection weights and combined,” wherein the examiner interprets weighting processed quantities by corresponding variable-selection weights and combining the weighted quantities to be the same as using a weighted sum because both multiply corresponding quantities by weights and aggregate the weighted quantities into a combined result.)
Zhu, Rangapuram, Lim, and the instant application are analogous art because they are all directed to deep-learning-based time-series prediction and the quantitative processing of features or uncertainty-related information used to generate or evaluate future-state predictions.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the computer-implemented time-series prediction apparatus disclosed by Zhu to configure its processor to include the deep recurrent state-space-model structure, temporal-feature encoding, forecast-time processing, and associated time-varying covariates disclosed by Rangapuram. One would be motivated to do so to effectively model temporal structure and contextual information over a defined forecast horizon and thereby improve the quality and robustness of future-state predictions, as suggested by Rangapuram ([Rangapuram, page 3, sec. 3] “SSMs model the temporal structure of the data via a latent state” AND [Rangapuram, page 2, sec. 3] “Our goal is to produce a set of probabilistic forecasts”).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to include the “weighted by their variable selection weights and combined” technique taught by Lim. One would be motivated to do so to efficiently aggregate multiple uncertainty-related quantities into a single instability measure while reducing the influence of less relevant information, as suggested by Lim ([Lim, page 1752, sec. 4.2] “variable selection also allows TFT to remove any unnecessary noisy inputs which could negatively impact performance”).
Claim(s) 2, 9, 12, and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Zhu in view of Rangapuram, further in view of Lim and NPL reference “Latent ODEs for Irregularly-Sampled Time Series”, by Rubanova et al. (referred herein as Rubanova).
Regarding claim 2, Zhu, Rangapuram, and Lim teaches The method according to claim 1, (see rejection of claim 1).
Zhu, Rangapuram, and Lim wherein the step of preprocessing past state data comprises: removing outliers from the past data; and calculating a time series length of the past data.
Rubanova teaches wherein the step of preprocessing past state data comprises: removing outliers from the past data; ([Rubanova, page 8, sec 4.4] “To speed up training, we rounded the observation times to the nearest minute, reducing the number of measurements only 2-fold” AND [Rubanova, page 9, sec 5], “used a binary mask to indicate the missing measurements and reported that RNNs performs better with zero-filling than with imputed values.” wherein the examiner interprets “zero-filling” and “rounded the observation times to the nearest minute, reducing the number of measurements” to be the same as removing outliers from the past data because they are both directed to a preprocessing step that filters or reduces raw data points to create a cleaner dataset for model training.)
and calculating a time series length of the past data. ([Rubanova, page 8, sec 4.4] “Hence, there are still 2880 (6048) possible measurement times per time series under our model's preprocessing, while the previous standard was to used only 48 possible measurement times…We evaluated our model on the PhysioNet Challenge 2012 dataset [Silva et al., 2012], which contains 8000 time series, each containing measurements from the first 48 hours”, wherein the examiner interprets “2880 (6048) possible measurement times per time series” and “8000 time series, each containing measurements from the first 48 hours” to be the same as calculating a time series length of the past data because they are both directed to determining and tracking the number of temporal observations or time points that constitute each time series sequence.)
Zhu, Rangapuram, Lim, Rubanova, and the instant application are analogous art because they are all directed to preprocessing time-series data for use in machine-learning-based prediction models.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the method claim 1 disclosed by Zhu, Rangapuram, and Lim to include the rounding to nearest times preprocessing technique disclosed by Rubanova. One would be motivated to do so to efficiently reduce noise and irregularities in historical time-series inputs, thereby improving model training robustness and computational efficiency, as suggested by Rubanova ([Rubanova, page 8, sec 4.4] “To speed up training, we rounded the observation times to the nearest minute, reducing the number of measurements only 2-fold.”).
Regarding claim 9, Zhu, Rangapuram, and Lim teach The method according to claim 1, (see rejection of claim 1).
Zhu, Rangapuram, and Lim do not teach wherein the step of predicting a future state comprises: calculating a prediction point-in-time by receiving a future point-in-time that a user wants to know; andpredicting a future state of the user through execution of an algorithm based on the prediction point-in-time and the trained model.
Rubanova teaches wherein the step of predicting a future state comprises: calculating a prediction point-in-time by receiving a future point-in-time that a user wants to know; ([Rubanova, page 2, sec 2] “The hidden state h(t) is defined at all times, and can be evaluated at any desired times using a numerical ODE solver” AND [Rubanova, page 5-6, sec 4] “Latent ODEs can often reconstruct trajectories reasonably well given a small subset of points, and provide an estimate of uncertainty over both the latent trajectories and trajectory (100 points) from a subset of 30 points. At test time, we conditioned this model on a subset of 10, 30 or 50 points…predict points on [2.5; 5] interval (blue area). A Latent ODE with an ODE-RNN encoder was able to extrapolate the time series far beyond the training interval and maintain periodic dynamics”, wherein the examiner interprets “evaluated at any desired times” and “Latent ODEs can often reconstruct trajectories reasonably well given a small subset of points, and provide an estimate of uncertainty over both the latent trajectories…was able to extrapolate the time series far beyond the training interval and maintain periodic dynamics” to be the same as calculating a prediction point-in-time by receiving a future point-in-time that a user wants to know because they are both directed to accepting a specified target time point at which the user desires to obtain a prediction.)
and predicting a future state of the user through execution of an algorithm based on the prediction point-in-time and the trained model. ([Rubanova, page 2, sec 2] “The hidden state h(t) is defined at all times, and can be evaluated at any desired times using a numerical ODE solver: h0, . . . , hN = ODESolve(fθ, h0,(t0, . . . , tN ))” AND [Rubanova, page 3, sec 3.2] “the generative model is defined by ODE whose initial latent state z0 determines the entire trajectoryz0, z1, . . . , zN = ODESolve(fθ, z0,(t0, t1, . . . , tN ))” AND [Rubanova, page 6] “At test time, we conditioned this model on a subset of 10, 30 or 50 points”, wherein the examiner interprets “ODESolve(fθ, z0,(t0, t1, . . . , tN ))” and “conditioned this model on a subset of 10, 30 or 50 points” to be the same as predicting a future state of the user through execution of an algorithm based on the prediction point-in-time and the trained model because they are both directed to using a trained model with an algorithmic solver to generate predictions at the specified future time points.)
Zhu, Rangapuram, Lim, Rubanova, and the instant application are analogous art because they are all directed to methods for predicting a future state from time series data at a specified future time point using trained models and algorithmic execution.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the method claim 1 disclosed by Zhu, Rangapuram, and Lim to include the capability of “evaluated at any desired times” disclosed by Rubanova. One would be motivated to do so to flexibly generate predictions at arbitrary future time points selected by a user, as suggested by Rubanova ([Rubanova, page 2, sec 2], “The hidden state h(t) is defined at all times, and can be evaluated at any desired times using a numerical ODE solver.”).
Regarding claim 10, Zhu, Rangapuram, Lim, and Rubanova teach The method according to claim 9, (see rejection of claim 9).
Zhu further teaches calculating at least one of reliability of the future state, …. ([Zhu, page 3, sec 3] “Specifically, we would like to quantify the prediction standard error, η, so that an approximate α-level prediction interval can be constructed” AND [Zhu, page 1, sec 1] “assessing how much to trust the forecast produced by the model”, wherein the examiner interprets “quantify the prediction standard error”, “trust the forecast”, and “prediction interval” to be the same as “reliability of the future state” because they are all directed to computing a quantitative measure indicating how dependable the predicted future outcome is).
and instability ([Zhu, page 3, sec 3.1] “the variance of the prediction distribution quantifies the prediction uncertainty”, wherein the examiner interprets “variance” and “prediction uncertainty” to be the same as “instability” because they are both directed to variability/volatility in the predicted future outcome).
Lim further teaches:
wherein the step of predicting the future state comprises: ([Lim, page 1750, sec 2] “direct methods are trained to explicitly generate forecasts”, wherein the examiner interprets “explicitly generate forecasts” to be the same as “predicting the future state” because they are both directed to producing a predicted future outcome from time series inputs).
a prediction basis of the future state, ([Lim, page 1749, sec 1] “TFT enables three valuable interpretability use cases: helping users identify (i) globally-important variables for the prediction problem” AND [Lim, page 1751, sec 3] “Fig. 2. TFT architecture. TFT inputs static metadata, time-varying past inputs and time-varying a priori known future inputs”, wherein the examiner interprets “interpretability use cases” and “globally-important variables”, as well as “time-varying a priori known future inputs” to be the same as “a prediction basis of the future state” because they are both directed to providing an explanation for why the model produced the future prediction).
Zhu, Rangapuram, Lim, Rubanova, and the instant application are analogous art because they are all directed to predicting a future state from time series data.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the method claim 9 disclosed by Zhu, Rangapuram, Lim, and Rubanova to include the interpretability technique disclosed by Lim. One would be motivated to do so to effectively provide explanatory information that clarifies how and why a predicted future state was generated, as suggested by Lim ([Lim, page 1749, sec 1] “interpretability use cases: helping users identify … globally-important variables for the prediction problem.”).
Regarding claim 12, Zhu, Rangapuram, and Lim teaches The apparatus according to claim 11, (see rejection of claim 11).
Zhu, Rangapuram, and Lim wherein the time series feature preprocessing unit comprises: an outlier-processing device removing outliers from the past data; and a time series length calculator calculating a time series length of the past data.
Rubanova teaches:
wherein the time series feature preprocessing unit comprises: an outlier-processing device removing outliers from the past data; ([Rubanova, page 8, sec 4.4] “To speed up training, we rounded the observation times to the nearest minute, reducing the number of measurements only 2-fold” AND [Rubanova, page 9, sec 5] “used a binary mask to indicate the missing measurements and reported that RNNs performs better with zero-filling than with imputed values”, wherein the examiner interprets “zero-filling” and “rounded the observation times to the nearest minute, reducing the number of measurements” to be the same as an outlier-processing device removing outliers from the past data because they are both directed to a preprocessing component that filters or reduces raw data points to create a cleaner dataset for model training.)
and a time series length calculator calculating a time series length of the past data. ([Rubanova, page 8, sec 4.4] “Hence, there are still 2880 (6048) possible measurement times per time series under our model's preprocessing, while the previous standard was to used only 48 possible measurement times…We evaluated our model on the PhysioNet Challenge 2012 dataset [Silva et al., 2012], which contains 8000 time series, each containing measurements from the first 48 hours”, wherein the examiner interprets “2880 (6048) possible measurement times per time series” and “8000 time series, each containing measurements from the first 48 hours” to be the same as a time series length calculator calculating a time series length of the past data because they are both directed to a computational element that determines and tracks the number of temporal observations or time points that constitute each time series sequence.)
Zhu, Rangapuram, Lim, Rubanova, and the instant application are analogous art because they are all directed to preprocessing time-series data for use in machine-learning-based prediction models.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the method claim 11 disclosed by Zhu, Rangapuram, and Lim to include the rounding to nearest times preprocessing technique disclosed by Rubanova. One would be motivated to do so to efficiently reduce noise and irregularities in historical time-series inputs, thereby improving model training robustness and computational efficiency, as suggested by Rubanova ([Rubanova, page 8, sec 4.4] “To speed up training, we rounded the observation times to the nearest minute, reducing the number of measurements only 2-fold.”).
Regarding claim 19, Zhu, Rangapuram, and Lim teach The apparatus according to claim 11, (see rejection of claim 11).
Zhu, Rangapuram, and Lim wherein the future state prediction unit calculates a prediction point-in-time by receiving a future point-in-time that a user wants to know, and predicts a future state of a user through execution of an algorithm based on the prediction point-in-time and the trained model.
Rubanova teaches wherein the future state prediction unit calculates a prediction point-in-time by receiving a future point-in-time that a user wants to know, ([Rubanova, page 2, sec 2] “The hidden state h(t) is defined at all times, and can be evaluated at any desired times using a numerical ODE solver” AND [Rubanova, page 5-6, sec 4] “Latent ODEs can often reconstruct trajectories reasonably well given a small subset of points, and provide an estimate of uncertainty over both the latent trajectories and trajectory (100 points) from a subset of 30 points. At test time, we conditioned this model on a subset of 10, 30 or 50 points…predict points on [2.5; 5] interval (blue area). A Latent ODE with an ODE-RNN encoder was able to extrapolate the time series far beyond the training interval and maintain periodic dynamics”, wherein the examiner interprets “evaluated at any desired times” and “Latent ODEs can often reconstruct trajectories reasonably well given a small subset of points, and provide an estimate of uncertainty over both the latent trajectories…was able to extrapolate the time series far beyond the training interval and maintain periodic dynamics” to be the same as calculating a prediction point-in-time by receiving a future point-in-time that a user wants to know because they are both directed to accepting a specified target time point at which the user desires to obtain a prediction.”)
and predicts a future state of a user through execution of an algorithm based on the prediction point-in-time and the trained model. ([Rubanova, page 2, sec 2] “The hidden state h(t) is defined at all times, and can be evaluated at any desired times using a numerical ODE solver: h0, . . . , hN = ODESolve(fθ, h0,(t0, . . . , tN ))” AND [Rubanova, page 3, sec 3.2] “the generative model is defined by ODE whose initial latent state z0 determines the entire trajectoryz0, z1, . . . , zN = ODESolve(fθ, z0,(t0, t1, . . . , tN ))” AND [Rubanova, page 6] “At test time, we conditioned this model on a subset of 10, 30 or 50 points”, wherein the examiner interprets “ODESolve(fθ, z0,(t0, t1, . . . , tN ))” and “conditioned this model on a subset of 10, 30 or 50 points” to be the same as predicts a future state of a user through execution of an algorithm based on the prediction point-in-time and the trained model because they are both directed to using a trained model with an algorithmic solver to generate predictions at the specified future time points.)
Zhu, Rangapuram, Lim, Rubanova, and the instant application are analogous art because they are all directed to methods for predicting a future state from time series data at a specified future time point using trained models and algorithmic execution.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the method claim 11 disclosed by Zhu and Rangapuram to include the capability of “evaluated at any desired times” disclosed by Rubanova. One would be motivated to do so to flexibly generate predictions at arbitrary future time points selected by a user, as suggested by Rubanova ([Rubanova, page 2, sec 2], “The hidden state h(t) is defined at all times, and can be evaluated at any desired times using a numerical ODE solver.”).
Claim(s) 6 and 16 are rejected under 35 U.S.C. 103 as being unpatentable over Zhu in view of Rangapuram, further in view of Lim and NPL reference “Mixture Density Networks”, by Bishop et al. (referred herein as Bishop).
Regarding claim 6, Zhu, Rangapuram, and Lim teach The method according to claim 1, (see rejection of claim 1).
Zhu, Rangapuram, and Lim do not teach further comprising: applying a suitable feature to the preprocessed past state data by modeling the environment condition feature; creating a complexity distribution of the preprocessed past state data; creating a complexity distribution sample based on the created complexity distribution; and creating future state data based on the created complexity distribution sampling.
Bishop teaches:
further comprising: applying a suitable feature to the preprocessed past state data by modeling the environment condition feature; creating a complexity distribution of the preprocessed past state data; creating a complexity distribution sample based on the created complexity distribution; and creating future state data based on the created complexity distribution sampling.
further comprising: applying the suitable feature to the preprocessed past state data by modeling the environment condition feature; ([Bishop, page 7, sec 3] “we now take the various parameters of the mixture model, namely the mixing coefficients αi(x), the means μi(x) and the variances σi(x), to be general (continuous) functions of x…This is achieved by modelling them using the outputs of a conventional neural network which takes x as its input”, wherein the examiner interprets “parameters of the mixture model…modelling them using the outputs of a conventional neural network which takes x as its input” to be the same as applying the suitable feature to the preprocessed past state data by modeling the environment condition feature because they are both directed to incorporating input-dependent contextual information that influences the output distribution parameters.)
creating a complexity distribution of the preprocessed past state data; ([Bishop, page 6, sec 3] “The probability density of the target data is then represented as a linear combination of kernel functions in the form p(t | x) = Σαi(x)φi(t | x)…we shall restrict attention to kernel functions which are Gaussian of the form φi(t | x) = 1/(2π)^(c/2)σi(x)^c exp{-||t - μi(x)||^2 / 2σi(x)^2}”, wherein the examiner interprets “probability density of the target data is then represented as a linear combination of kernel functions” and “kernel functions which are Gaussian” to be the same as creating a complexity distribution of the preprocessed past state data because they are both directed to constructing a mixture of multiple probability distributions that can model complex, multi-modal data patterns.)
creating a complexity distribution sample based on the created complexity distribution; ([Bishop, page 12, sec 3] “the most probable branch of the solution, assuming the components are well separated and have negligible overlap, is given by max{αi(x)}” AND [Bishop, page 12, sec 3] “The required value of t is then given by the corresponding centre μi”, wherein the examiner interprets “most probable branch of the solution” and “required value of t is then given by the corresponding centre μi” to be the same as creating a complexity distribution sample based on the created complexity distribution because they are both directed to selecting or sampling a specific output value from the constructed mixture distribution.)
and creating future state data based on the created complexity distribution sampling. ([Bishop, page 11, sec 3] “One of the simplest statistics is the mean, corresponding to the conditional average of the target data, given by <t | x> = Σαi(x)μi(x)” AND [Bishop, page 18, sec 5] “Having obtained a good representation for the conditional density of the target data, it is then in principle straightforward to calculate any desired statistic from that distribution”, wherein the examiner interprets “conditional average of the target data, given by <t | x> = Σαi(x)μi(x)” and “calculate any desired statistic from that distribution” to be the same as creating future state data based on the created complexity distribution sampling because they are both directed to generating predicted output values by computing statistics or samples from the learned mixture distribution model.)
Zhu, Rangapuram, Lim, Bishop, and the instant application are analogous art because they are all directed to probabilistic time series forecasting methods that generate future state data.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the method claim 1 disclosed by Zhu, Rangapuram, and Lim to include the technique to perform a linear combination of kernel functions disclosed by Bishop. One would be motivated to do so to effectively model complex, multi-modal future outcome distributions, as suggested by Bishop ([Bishop, page 6, sec 3] “The probability density of the target data is then represented as a linear combination of kernel functions”).
Regarding claim 16, Zhu, Rangapuram, and Lim teach The apparatus according to claim 11, (see rejection of claim 1).
Zhu, Rangapuram, and Lim do not teach wherein the environment feature processing device creates a complexity distribution of the preprocessed past state data, a complexity distribution sample based on the created complexity distribution, and future state data based on the created complexity distribution sampling.
Bishop teaches:
wherein the environment feature processing device creates a complexity distribution of the preprocessed past state data, ([Bishop, page 6, sec 3] “The probability density of the target data is then represented as a linear combination of kernel functions in the form p(t | x) = Σαi(x)φi(t | x)” AND [Bishop, page 6, sec 3] “we shall restrict attention to kernel functions which are Gaussian of the form φi(t | x) = 1/(2π)^(c/2)σi(x)^c exp{-||t - μi(x)||^2 / 2σi(x)^2}”, wherein the examiner interprets “probability density of the target data is then represented as a linear combination of kernel functions” and “kernel functions which are Gaussian” to be the same as the environment feature processing device creates a complexity distribution of the preprocessed past state data because they are both directed to a processing component that constructs a mixture of multiple probability distributions that can model complex, multi-modal data patterns.)
a complexity distribution sample based on the created complexity distribution, ([Bishop, page 12, sec 3] “the most probable branch of the solution, assuming the components are well separated and have negligible overlap, is given by max{αi(x)}” AND [Bishop, page 12, sec 3] “The required value of t is then given by the corresponding centre μi”, wherein the examiner interprets “most probable branch of the solution” and “required value of t is then given by the corresponding centre μi” to be the same as a complexity distribution sample based on the created complexity distribution because they are both directed to selecting or sampling a specific output value from the constructed mixture distribution.)
and future state data based on the created complexity distribution sampling. ([Bishop, page 11, sec 3] “One of the simplest statistics is the mean, corresponding to the conditional average of the target data, given by <t | x> = Σαi(x)μi(x)” AND [Bishop, page 18, sec 5] “Having obtained a good representation for the conditional density of the target data, it is then in principle straightforward to calculate any desired statistic from that distribution”, wherein the examiner interprets “conditional average of the target data, given by <t | x> = Σαi(x)μi(x)” and “calculate any desired statistic from that distribution to be the same as future state data based on the created complexity distribution sampling because they are both directed to generating predicted output values by computing statistics or samples from the learned mixture distribution model.)
Zhu, Rangapuram, Lim, Bishop, and the instant application are analogous art because they are both directed to generating future state data for time series using probabilistic modeling of latent state dynamics, including learning distributions over future outcomes from historical observations.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the apparatus claim 11 disclosed by Zhu, Rangapuram, and Lim to include the “linear combination of kernel functions” disclosed by Bishop. One would be motivated to do so to effectively model complex and multi-modal predictive distributions for future state estimation, as suggested by Bishop ([Bishop, page 6, sec 3] “The probability density of the target data is then represented as a linear combination of kernel functions.”).
Conclusion
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 DEVAN KAPOOR whose telephone number is (703)756-1434. The examiner can normally be reached Monday - Friday: 9:00AM - 5:00 PM EST (times may vary).
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, David Yi can be reached at (571) 270-7519. 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.
/DEVAN KAPOOR/Examiner, Art Unit 2126
/DAVID YI/Supervisory Patent Examiner, Art Unit 2126
Prosecution Insights