DETAILED ACTION
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 .
This action is responsive to the Applicant’s amendments filed on 01/23/2026. Claims 1-16 and 19-20 remain pending in the application. Claims 1, 8, 10-16, and 19-20 have been amended. Claims 17-18 have been canceled. Any examiner’s note, objection, and rejection not repeated is withdrawn due to Applicant’s amendment.
Priority
This application is a National Stage entry under 35 U.S.C. 371 of International Application No. PCT/US2021/056258, filed 10/22/2021. The international filing date of 10/22/2021 is acknowledged as the U.S. filing date in accordance with 35 U.S.C. 363.
Benefit is claimed of U.S. Provisional Application No. 63/104,260 filed 10/22/2020 under 35 U.S.C. 119(e).
Information Disclosure Statement
The information disclosure statements (IDS) submitted on 05/15/2024, 05/15/2024, and 01/08/2025, and 01/16/2026 are in compliance with the provisions of 37 CFR 1.97. Accordingly, the information disclosure statements are being considered by the Examiner.
Examiner’s Note
The Examiner cites particular columns, paragraphs, figures, and line numbers in the references as applied to the claims below for the convenience of the applicant. Although the specified citations are representative of the teachings in the art and are applied to the specific limitations within the individual claim, other passages and figures may also apply. It is respectfully requested that, in preparing responses, the Applicant fully consider the references in its entirety as potentially teaching all or part of the claimed invention, as well as the context of the passage as taught by the prior art or disclosed by the Examiner.
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.
Claim 1 is rejected under 35 U.S.C. 103 as being unpatentable over Morris et al. (US 20220083378 A1) hereafter Morris in view of Higa (US 20220012540 A1), further in view of Matsuura et al. (US 20200034209 A1) hereafter Matsuura, further in view of Li et al. (US 20200409754 A1) hereafter Li, further in view of Kroeger et al. (US 20140142998 A1) hereafter Kroeger.
Regarding claim 1, Morris teaches:
A method for runtime task scheduling in a heterogeneous multi-core computing system, the method comprising:
obtaining an application comprising a plurality of tasks (Paragraphs 53 and 136; “The compiler 120 may decompose the DL application 110 into tasks”, where “The controller 710 may receive metadata of each of tasks from a host”, clearly discloses a controller may obtain a plurality of tasks that in combination, correspond to an application);
obtaining policies for task scheduling (Paragraphs 74-77; “The policy generator 410 may execute an actor network 411 based on the input vector 451”, in which “the actor network 411 may also be referred to as a ‘policy network’”, and “the policy-based scheduler 420 may perform first resource scheduling on the task 1 210 and the task A based on the determined action”. The cited paragraphs disclose a policy generator 410 executing a policy network which generates scheduling policies through a DNN.);
and scheduling the plurality of tasks on a heterogeneous set of processing elements according to the policies (Paragraph 71, 76-77; “The policy-based scheduler 420 may select the task a and the task 1 210 as tasks to be executed in parallel”. Further, “The determined action may include a priority of each of hardware configurations T 1[A], T 1[B] and T 1[C] for the task 1 210, and a priority of each of hardware configurations for the task a. The actor network 411 may transmit the determined action to the policy-based scheduler 420”, then “The policy-based scheduler 420 may perform first resource scheduling on the task 1 210 and the task a based on the determined action”. In other words, the policy-based scheduler 420 applies the obtained policies to schedule multiple tasks. The scheduling is performed across multiple hardware configurations which correspond to a heterogeneous set of processing elements, based on the determined action of the policy network, corresponding to being based on the policies);
wherein the processing elements are arranged in processing clusters (Paragraph 51; “The resources 140 may include a plurality of clusters (for example, neural processing unit (NPU) clusters)”, NPU cluster corresponding to a heterogeneous set of processing elements arranged in a cluster.).
Morris does not teach imitation learning.
However, Higa teaches:
imitation learning (Paragraph 25; “The function that outputs an action to be taken by the agent according to the state of the target environment is called a policy. The imitation learning unit 30, described below, generates a policy through imitation learning.”).
Morris and Higa are considered to be analogous to the claimed invention because they are in the same field of resource management using ML. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Morris to incorporate the teachings of Higa and have utilized the DNN system of Morris with IL taught by Higa. A person of ordinary skill in the art would have recognized this utilization as a known technique and have been motivated by the disclosure in Higa, “by performing parameter tuning on this reward function, it becomes possible to obtain a highly accurate solution” as disclosed in Paragraph 54 of Higa.
Morris in view of Higa does not teach a first-level policy for selecting one of the processing clusters; a second-level policy for selecting a processing element within the one selected processing cluster for each of the plurality of tasks; the scheduling of a task of the plurality of tasks on the heterogeneous set of processing elements is based on: an average execution time of the task on the processing element; a relative order of the task in the task queue; a total number of tasks; precedence constraints between the tasks; readiness of the processing elements to execute tasks.
However, Matsuura teaches:
a first-level policy for selecting one of the processing clusters (Paragraph 87; “a cluster that is less affected by the coexistence of the execution of the application is selected from the execution situation of the application in each of the clusters stored in the data table 400 that is managed by the application progress information storage unit 122”, which discloses selecting a cluster from a plurality of clusters based on execution conditions. This corresponds to a first-level decision determining which cluster executes the application, performed according to evaluation criteria derived from the situation of the application from the data table, which corresponds to a policy for selecting among alternatives based on observed conditions.);
a second-level policy for selecting a processing element within the one selected processing cluster for each of the plurality of tasks (Paragraph 90; “the processing server 130 to be allocated from the selected cluster is selected from the data table 300 that is managed by the resource allocation information storage unit 121”, which selects a processing server from a selected cluster based on resource allocation info. The processing server corresponds to a processing element within the clusters which comprise multiple processing servers. Each processing server is a unit of computation to which tasks are assigned. The selection of the processing server from within the selected cluster corresponds to a second-level policy-based selection of a processing element within the processing cluster. The claim defines processing elements functionally as entities being scheduled and arranged into clusters. Matsuura teaches processing servers that are scheduled and that clusters are composed of servers. Therefore, the processing servers of Matsuura satisfy the claimed processing elements.).
Morris, Higa, and Matsuura are considered to be analogous to the claimed invention because they are in the same field of resource management. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Morris in view of Higa to incorporate the teachings of Matsuura and combine the teachings of IL with the policy-based selection of a cluster and a particular processing element within the cluster of Matsuura applied on each of the plurality of tasks of Morris. Morris in view of Higa teaches implementing the use of imitation learning for resource management based on system conditions. Matsuura further teaches a hierarchical resource framework in which a cluster is selected and a processing server within the server is selected based on resource allocation information (Paragraphs 87, 90), thereby decomposing resource selection into multiple levels of decision-making. It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to have incorporated the hierarchical selection framework of Matsuura into the imitation learning-based system of Morris in view of Higa by applying the IL policies to each level of the hierarchical selection, because Matsuura’s decomposition of resource allocation into cluster-level and intra-cluster selection is a known method in the art yielding the predictable result of reducing decision complexity and improving scalability in distributed systems. Higa further teaches that such decision-making processes may be implemented as imitation learning processes.
Morris in view of Higa, further in view of Matsuura does not teach the scheduling of a task of the plurality of tasks on the heterogeneous set of processing elements is based on: an average execution time of the task on the processing element; a relative order of the task in the task queue; a total number of tasks; precedence constraints between the tasks; readiness of the processing elements to execute tasks.
However, Li teaches:
the scheduling of a task of the plurality of tasks on the heterogeneous set of processing elements is based on (Paragraph 15; “Based on the collected performance metrics, the server computer system updates a scheduling algorithm for assigning program tasks to queues in the first and second sets of task queues. The server computer system receives program tasks and schedules them into particular queues based on the updated algorithm.”):
an average execution time of the task on the processing element (Paragraph 42; “Scheduling optimizer 106 retrieves at least a portion, as indicated by reference 126, of collected performance metrics 104 and may use this information to update information regarding performance of task execution and workflow efficiency within server computer system 100. Scheduling optimizer 106 may track, for example, an average execution time for program tasks assigned to the first set of task queues (avg. execution time for first queue tasks) 350”);
a relative order of the task in the task queue (Paragraph 18; “With an in-order task queue, if the next program task in the task queue utilizes the particular resource or memory, then the processor core stalls until the given program task completes and the particular resource or memory is available.”);
a total number of tasks (Paragraph 22; “Various criteria may be used to select a particular task queue for a given program task, such as a particular job associated with the given program task, a priority level associated with the given program task, a current number of program tasks assigned to each task queue”. Paragraph 45 further discloses “Estimated time to execute all tasks in each first queue 358 and estimated time to execute all tasks in each second queue 360, are summations of the estimated execution times for each program task assigned and currently pending in each of the task queues in the first and second sets of task queues 110 and 112”. It would have been obvious to a person of ordinary skill in the art to determine a total number of tasks by applying the same aggregation technique of summation across tasks to the task counts, representing a straightforward application of an explicitly disclosed aggregation operation to an analogous metric, yielding the predictable result of a total task count for the system.);
Morris, Higa, Matsuura, and Li are considered to be analogous to the claimed invention because they are in the same field of resource management. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Morris in view of Higa, further in view of Matsuura to incorporate the teachings of Li and have scheduled tasks on processing elements based on an average execution time, a relative order in the queue, and a total number of tasks. A person of ordinary skill in the art would have been motived to utilize related and derived workload metrics such as average execution time, total number of tasks, and queue order, in scheduling decisions, as these metrics provide normalized indicators of system load. Applying such metrics to the task scheduling methodology of Morris in view of Higa further in view of Matsuura would be a straightforward extension of Li’s metric-based scheduling approach, yielding the predictable result of allocating tasks based on workload characteristics.
Morris in view of Higa, further in view of Matsuura, further in view of Li does not teach precedence constraints between the tasks; readiness of the processing elements to execute tasks.
However, Kroeger teaches:
precedence constraints between the tasks (Paragraph 52; “tasks in a story can not only be used to approximate the amount of resources required to complete the story, but also show how the story should be completed by defining dependencies, precedence or any type of relationships among the tasks.”, explicitly discloses defining precedence among tasks.);
readiness of the processing elements to execute tasks (Paragraph 46; “management system 200 can proceed to perform task assignment according to the plan (step 140) by assigning a task to a resource (e.g., a team member) for execution as soon as the resource becomes available”, where the availability of the resource indicates that it is eligible and capable of executing a task, thereby corresponding to readiness to execute tasks.).
Morris, Higa, Matsuura, Li, and Kroeger are considered to be analogous to the claimed invention because they are in the same field of resource management. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Morris in view of Higa, further in view of Matsuura, further in view of Li to incorporate the teachings of Kroeger and have scheduled tasks based on precedence constraints between tasks and readiness of processors to execute tasks. Morris in view of Higa, further in view of Matsuura, further in view of Li disclose scheduling and selection mechanisms. Kroeger teaches assigning tasks to available resources and enforcing task dependencies and precedence, thereby ensuring that task execution is feasible and consistent with ordering constraints. Incorporating these teachings into the existing scheduling framework would apply known constraint checking and optimization methods to the outputs of the existing IL-based hierarchical scheduling policies, which is a predictable use of prior art elements according to their established functions resulting in coordinated tasks election, resource allocation, and execution ordering when resources are available and precedence constraints are satisfied.
Claims 2-4 and 10 are rejected under 35 U.S.C. 103 as being unpatentable over Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger, further in view of Bogdan et al. (US 20210049465 A1) hereafter Bogdan.
Regarding claim 2, Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger teach the method of claim 1. Higa teaches:
imitation learning (Paragraph 25; “The function that outputs an action to be taken by the agent according to the state of the target environment is called a policy. The imitation learning unit 30, described below, generates a policy through imitation learning.”).
Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger does not teach wherein obtaining the IL policies comprises training the IL policies offline.
However, Bogdan teaches:
wherein obtaining the IL policies comprises training the IL policies offline (Paragraph 96; “NN Classifiers: In order to detect specialized features such as FFT and MM in applications, it is difficult for any partitioning scheme to cut the graph into several sub-graphs representing functionalities. Therefore, NN classifiers are adopted to first offline learn the structures of numerous features in a specific domain such as signal processing (FFT, compression) and deep learning (MM, SGD); and then identify the topology of the existing features given new applications.”).
Morris, Higa, Matsuura, Li, Kroeger, and Bogdan are considered to be analogous to the claimed invention because they are in the same field of task scheduling using ML. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger to incorporate the teachings of Bogdan and train the IL policies offline. A person of ordinary skill in the art would have recognized this as an obvious design decision for DNNs.
Regarding claim 3, Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger, further in view of Bogdan teach the method of claim 2. Bogdan teaches:
wherein training the IL policies offline uses supervised machine learning (Paragraph 77; “Another aspect of the present embodiment, is the proper training of the neural network classifiers. In this regard, the one or more trained neural network classifiers are trained by collecting a group of training applications with different features. Examples of such different features include, but are not limited to, loops for CPUs; loops for GPUs; neurons in a neural network, activation functions, matrix multiplications, vector multiplication, gradient descent, for-loops in a machine learning domain, fast Fourier transforms, matrix multiplication, and combinations thereof.” The use of gradient descent inherently teaches supervised learning because the loss function of gradient descent requires known labels.).
Regarding claim 4, Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger, further in view of Bogdan teach the method of claim 3. Bogdan teaches:
wherein the supervised machine learning comprises one or more of a linear regression, a regression tree, or a neural network (Paragraph 77; “Another aspect of the present embodiment, is the proper training of the neural network classifiers. In this regard, the one or more trained neural network classifiers are trained by collecting a group of training applications with different features. Examples of such different features include, but are not limited to, loops for CPUs; loops for GPUs; neurons in a neural network, activation functions, matrix multiplications, vector multiplication, gradient descent, for-loops in a machine learning domain, fast Fourier transforms, matrix multiplication, and combinations thereof” explicitly covers the neural network element of the claim.).
Regarding claim 10, Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger teach the method of claim 1. Morris teaches:
scheduling application tasks for multi-tasking across a plurality of applications on according to the policies (Paragraph 71; “the policy-based scheduler may select tasks to be executed in parallel” and “may perform rescheduling” explicitly discloses performing task scheduling. Further, “when a new DL application A is enqueued… the policy-based scheduler may determine that a task a of the DL application A and the task 1 210 of the DL application 110 are independent [and] may select the task a and the task 1 210 as tasks to be executed in parallel”, teaching scheduling across multiple applications and multi-tasking. Paragraphs 74-77 further discloses “the policy generator 410 may execute an actor network 411”, which “may determine or select an action for the input vector”, which causes the policy based scheduler to perform resource scheduling based on the determined action, corresponding to scheduling based on policies.).
Higa teaches:
IL (Paragraph 25; “The function that outputs an action to be taken by the agent according to the state of the target environment is called a policy. The imitation learning unit 30, described below, generates a policy through imitation learning.”).
Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger does not teach a heterogeneous set of processing elements.
However, Bogdan teaches:
A heterogeneous set of processing elements (Paragraph 125; “a heterogeneous platform consisting of 32 CPUs and 32 GPUs is modeled using MacSim plus special HWAs with NoC substrate using BookSim2” explicitly discloses a heterogeneous hardware platform including elements such as CPUs, GPUs, and hardware accelerators.).
Morris, Higa, Matsuura, Li, Kroeger, and Bogdan are considered to be analogous to the claimed invention because they are in the same field of task scheduling using ML. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger to incorporate the teachings of Bogdan and utilize a heterogeneous set of processing elements. A person of ordinary skill in the art would have recognized the need for multiple types of processing elements to best process particular types of tasks.
Claims 5-8, 11-12, 16, and 19-20 are rejected under 35 U.S.C. 103 as being unpatentable over Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger, further in view of Bogdan, further in view of Khanna (US 20220083900 A1).
Regarding claim 5, Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger, further in view of Bogdan teach the method of claim 3. Higa teaches:
IL (Paragraph 25; “The function that outputs an action to be taken by the agent according to the state of the target environment is called a policy. The imitation learning unit 30, described below, generates a policy through imitation learning.”).
Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger, further in view of Bogdan does not teach constructing an oracle; and training the policies using the oracle.
However, Khanna teaches:
constructing an oracle (Paragraph 55; “the system 102 may store feature vectors labeled through oracle identification…” explicitly discloses the use of oracle identification as a mechanism for labeling feature vectors. This is equivalent to constructing an oracle in the sense of preparing a supervisory source of truth to train policies.);
and training the IL policies using the oracle (Paragraph 55; “The labeled feature vectors can be used by a supervised machine learning system to train a machine learning model”. Since the labeled feature vectors originate from oracle identification, the training process therefore uses an oracle.).
Morris, Higa, Matsuura, Li, Kroeger, Bogdan, and Khanna are considered to be analogous to the claimed invention because they are in the same field of task management using ML. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger, further in view of Bogdan to incorporate the teachings of Khanna and construct an oracle to train IL policies on. A person of ordinary skill in the art would understand that supervised training with oracle-provided labels is consistent with imitation learning practices in the art.
Regarding claim 6, Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger, further in view of Bogdan, further in view of Khanna teach the method of claim 5. Higa teaches:
IL (Paragraph 25; “The function that outputs an action to be taken by the agent according to the state of the target environment is called a policy. The imitation learning unit 30, described below, generates a policy through imitation learning.”).
Bogdan teaches:
wherein obtaining the policies further comprises generating training data for the policies using a simulation of the heterogeneous multi-core computing system (Paragraph 102; “When new applications arrive, it is first transformed into IDGs. Based on the graphs and their metrics, input data for NNs is prepared, which is then driven into NNs” describes creating input data from simulated applications, transformed into intermediate dependency graphs, which corresponds to generating training data. “The first trained NN figures out the types of special features… The second NN determines the exact locations and types of features” correspond to the generated input data being used to train neural networks encompassing IL policies. Further, “During testing and simulation, unknown applications are used” to test the validity of the solution. The system explicitly utilizes simulation with machine learning applications from the same domain as training workloads to generate input data. Because the applications exercise multiple heterogeneous resources, this constitutes simulation of a heterogeneous multi-core computing system for training.).
Regarding claim 7, Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger, further in view of Bogdan, further in view of Khanna teach the method of claim 6. Higa teaches:
IL (Paragraph 25; “The function that outputs an action to be taken by the agent according to the state of the target environment is called a policy. The imitation learning unit 30, described below, generates a policy through imitation learning.”).
Bogdan teaches:
wherein obtaining the policies further comprises improving the policies based on aggregated data from oracle actions and results of the policies during simulation (Paragraphs 124-125; “Trace-driven timing simulators… execute these instructions and return the next state and rewards back to SOSPCS to help agents optimize themselves”, in which the simulators generate results in the form of state and reward signals which are used by the RL/IL agent to improve policies. “The experimental results are compared with state-of-the-art HETS scheduling”, and “The baseline for the results is to run applications with parallel execution using METIS in the simulator” where SOSPCS aggregates and compares oracle/baseline scheduling actions defined in HETS with the results. The benchmarking serves as oracle actions, providing labeled data for policy improvement. The baseline scheduling algorithms act as oracles whose actions establish a reference policy. Returning the next state and rewards back to help agents optimize correspond to producing scheduling decisions whose performance outcomes, the states and rewards, are fed back into the agent. The use of a simulator explicitly discloses that these results are derived during simulation.).
Regarding claim 8, Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger, further in view of Bogdan, further in view of Khanna teach the method of claim 7. Higa teaches:
IL (Paragraph 25; “The function that outputs an action to be taken by the agent according to the state of the target environment is called a policy. The imitation learning unit 30, described below, generates a policy through imitation learning.”).
Khanna teaches:
labeling a current oracle action for a task (Paragraph 55; “the system 102 may store feature vectors labeled through oracle identification or through an inductive learning process in the training data repository 116. The labeled feature vectors can be used by a supervised machine learning system to train a machine learning model” explicitly shows labeling via oracle identification, where the oracle provides the correct output for each feature vector, analogous to labeling an oracle action for a given system state).
Bogdan teaches:
policy actions different from the oracle actions (Paragraphs 124-125; “The experimental results are compared with… HETS scheduling, and METIS graph partitioning approaches. The baseline for the results is to run applications with parallel execution using METIS in the simulator.” METIS and HETS provide baseline oracle actions. When SOSPCS schedules differently, the comparison identifies which action is the oracle and which is the policy action, corresponding to labeling the oracle action for a given task. The comparison of experimental results with METIS and HETS explicitly considers situations where the policy actions differ from the baseline scheduler.);
and retraining the IL policies using the aggregated data comprising the labeled oracle action and a corresponding system state (Paragraphs 124-125; “Trace-drive timing simulators… return the next state and rewards back to SOSPCS to help agents optimize themselves” discloses the agent updates, thereby retraining, its scheduling policy based on the simulator-provided feedback. Further, “Trace-driven timing simulators… execute these instructions and return the next state and rewards back” disclose baseline schedulers, corresponding to oracle actions, RL scheduling, corresponding to policy results, and system state/rewards from simulators. In combination, these elements make up the aggregated data used to train the agent.).
Regarding claim 11, Morris teaches:
A method comprising:
generating policies for task scheduling (Paragraph 74-77; “The policy generator 410 may execute an actor network 411” and “may determine or select an action”. Further, “the policy-based scheduler 420 may perform first resource scheduling on the task 1 210 and the task a based on the determined action”, which discloses a policy generator producing policy network outputs, corresponding to imitation learning outputs, applied to task scheduling.);
a heterogeneous set of processing elements (Paragraph 51; “The resources 140 may include a plurality of clusters (for example, neural processing unit (NPU) clusters)”, NPU cluster corresponding to a heterogeneous set of processing elements.);
the processing elements are arranged in processing clusters (Paragraph 51; “The resources 140 may include a plurality of clusters (for example, neural processing unit (NPU) clusters)”, NPU cluster corresponding to a heterogeneous set of processing elements arranged in a cluster.).
Morris does not teach a simulating a plurality of scheduling algorithms for a plurality of application tasks; selecting, by an oracle, actions for task scheduling; imitation learning (IL); or during runtime on a heterogeneous system on a chip (SoC), wherein the generating of the IL policies comprises training the IL policies using the oracle.
However, Higa teaches:
imitation learning (IL) (Paragraph 25; “The function that outputs an action to be taken by the agent according to the state of the target environment is called a policy. The imitation learning unit 30, described below, generates a policy through imitation learning.”).
Morris and Higa are considered to be analogous to the claimed invention because they are in the same field of resource management using ML. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Morris to incorporate the teachings of Higa and have utilized the DNN system of Morris with IL taught by Higa. A person of ordinary skill in the art would have recognized this utilization as a known technique and have been motivated by the disclosure in Higa, “by performing parameter tuning on this reward function, it becomes possible to obtain a highly accurate solution” as disclosed in Paragraph 54 of Higa.
Morris in view of Higa does not teach simulating a plurality of scheduling algorithms for a plurality of application tasks; selecting, by an oracle, actions for task scheduling; or during runtime on a heterogeneous system on a chip (SoC), wherein the generating of the IL policies comprises training the IL policies using the oracle; a first-level policy selects one of the processing clusters; a second-level policy selects a processing element within the one selected processing cluster to be scheduled for each of the plurality of tasks; and the task scheduling during runtime comprises scheduling a task of the plurality of tasks on the heterogeneous set of processing elements based on: an average execution time of the task on the processing element; a relative order of the task in a task queue; a total number of tasks; precedence constraints between the tasks; and readiness of the processing element to execute tasks.
However, Matsuura teaches:
a first-level policy for selecting one of the processing clusters (Paragraph 87; “a cluster that is less affected by the coexistence of the execution of the application is selected from the execution situation of the application in each of the clusters stored in the data table 400 that is managed by the application progress information storage unit 122”, which discloses selecting a cluster from a plurality of clusters based on execution conditions. This corresponds to a first-level decision determining which cluster executes the application, performed according to evaluation criteria derived from the situation of the application from the data table, which corresponds to a policy for selecting among alternatives based on observed conditions.);
a second-level policy for selecting a processing element within the one selected processing cluster for each of the plurality of tasks (Paragraph 90; “the processing server 130 to be allocated from the selected cluster is selected from the data table 300 that is managed by the resource allocation information storage unit 121”, which selects a processing server from a selected cluster based on resource allocation info. The processing server corresponds to a processing element within the clusters which comprise multiple processing servers. Each processing server is a unit of computation to which tasks are assigned. The selection of the processing server from within the selected cluster corresponds to a second-level policy-based selection of a processing element within the processing cluster. The claim defines processing elements functionally as entities being scheduled and arranged into clusters. Matsuura teaches processing servers that are scheduled and that clusters are composed of servers. Therefore, the processing servers of Matsuura satisfy the claimed processing elements.).
Morris, Higa, and Matsuura are considered to be analogous to the claimed invention because they are in the same field of resource management. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Morris in view of Higa to incorporate the teachings of Matsuura and combine the teachings of IL with the policy-based selection of a cluster and a particular processing element within the cluster of Matsuura applied on each of the plurality of tasks of Morris. Morris in view of Higa teaches implementing the use of imitation learning for resource management based on system conditions. Matsuura further teaches a hierarchical resource framework in which a cluster is selected and a processing server within the server is selected based on resource allocation information (Paragraphs 87, 90), thereby decomposing resource selection into multiple levels of decision-making. It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to have incorporated the hierarchical selection framework of Matsuura into the imitation learning-based system of Morris in view of Higa by applying the IL policies to each level of the hierarchical selection, because Matsuura’s decomposition of resource allocation into cluster-level and intra-cluster selection is a known method in the art yielding the predictable result of reducing decision complexity and improving scalability in distributed systems. Higa further teaches that such decision-making processes may be implemented as imitation learning processes.
Morris in view of Higa, further in view of Matsuura does not teach the scheduling of a task of the plurality of tasks is based on: an average execution time of the task on the processing element; a relative order of the task in the task queue; a total number of tasks; precedence constraints between the tasks; readiness of the processing elements to execute tasks.
However, Li teaches:
the scheduling of a task of the plurality of tasks is based on (Paragraph 15; “Based on the collected performance metrics, the server computer system updates a scheduling algorithm for assigning program tasks to queues in the first and second sets of task queues. The server computer system receives program tasks and schedules them into particular queues based on the updated algorithm.”):
an average execution time of the task on the processing element (Paragraph 42; “Scheduling optimizer 106 retrieves at least a portion, as indicated by reference 126, of collected performance metrics 104 and may use this information to update information regarding performance of task execution and workflow efficiency within server computer system 100. Scheduling optimizer 106 may track, for example, an average execution time for program tasks assigned to the first set of task queues (avg. execution time for first queue tasks) 350”);
a relative order of the task in the task queue (Paragraph 18; “With an in-order task queue, if the next program task in the task queue utilizes the particular resource or memory, then the processor core stalls until the given program task completes and the particular resource or memory is available.”);
a total number of tasks (Paragraph 22; “Various criteria may be used to select a particular task queue for a given program task, such as a particular job associated with the given program task, a priority level associated with the given program task, a current number of program tasks assigned to each task queue”. Paragraph 45 further discloses “Estimated time to execute all tasks in each first queue 358 and estimated time to execute all tasks in each second queue 360, are summations of the estimated execution times for each program task assigned and currently pending in each of the task queues in the first and second sets of task queues 110 and 112”. It would have been obvious to a person of ordinary skill in the art to determine a total number of tasks by applying the same aggregation technique of summation across tasks to the task counts, representing a straightforward application of an explicitly disclosed aggregation operation to an analogous metric, yielding the predictable result of a total task count for the system.);
Morris, Higa, Matsuura, and Li are considered to be analogous to the claimed invention because they are in the same field of resource management. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Morris in view of Higa, further in view of Matsuura to incorporate the teachings of Li and have scheduled tasks on processing elements based on an average execution time, a relative order in the queue, and a total number of tasks. A person of ordinary skill in the art would have been motived to utilize related and derived workload metrics such as average execution time, total number of tasks, and queue order, in scheduling decisions, as these metrics provide normalized indicators of system load. Applying such metrics to the task scheduling methodology of Morris in view of Higa further in view of Matsuura would be a straightforward extension of Li’s metric-based scheduling approach, yielding the predictable result of allocating tasks based on workload characteristics.
Morris in view of Higa, further in view of Matsuura, further in view of Li does not teach precedence constraints between the tasks; readiness of the processing elements to execute tasks.
However, Kroeger teaches:
precedence constraints between the tasks (Paragraph 52; “tasks in a story can not only be used to approximate the amount of resources required to complete the story, but also show how the story should be completed by defining dependencies, precedence or any type of relationships among the tasks.”, explicitly discloses defining precedence among tasks.);
readiness of the processing elements to execute tasks (Paragraph 46; “management system 200 can proceed to perform task assignment according to the plan (step 140) by assigning a task to a resource (e.g., a team member) for execution as soon as the resource becomes available”, where the availability of the resource indicates that it is eligible and capable of executing a task, thereby corresponding to readiness to execute tasks.).
Morris, Higa, Matsuura, Li, and Kroeger are considered to be analogous to the claimed invention because they are in the same field of resource management. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Morris in view of Higa, further in view of Matsuura, further in view of Li to incorporate the teachings of Kroeger and have scheduled tasks based on precedence constraints between tasks and readiness of processors to execute tasks. Morris in view of Higa, further in view of Matsuura, further in view of Li disclose scheduling and selection mechanisms. Kroeger teaches assigning tasks to available resources and enforcing task dependencies and precedence, thereby ensuring that task execution is feasible and consistent with ordering constraints. Incorporating these teachings into the existing scheduling framework would apply known constraint checking and optimization methods to the outputs of the existing IL-based hierarchical scheduling policies, which is a predictable use of prior art elements according to their established functions resulting in coordinated tasks election, resource allocation, and execution ordering when resources are available and precedence constraints are satisfied.
Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger does not teach simulating a plurality of scheduling algorithms for a plurality of application tasks; during runtime on a heterogeneous SoC, wherein the policies are trained using the SoC simulator; selecting, by an oracle, actions for task scheduling
However, Bogdan teaches:
simulating a plurality of scheduling algorithms for a plurality of application tasks (Paragraphs 124-125; “a heterogeneous platform consisting of 32 CPUs and 32 GPUs is modeled using Macsim plus special HWAs with NoC substrate using BookSim2” discloses a simulator for heterogeneous SoC systems with CPUs, GPUs, and HWAs. “The experimental results are compared with state-of-the-art HETS scheduling [26] and METIS graph partitioning [18] approaches” shows that the simulator is configured to run and compare multiple scheduling algorithms across multiple application tasks. “IR traces of tasks and PE indices to which these tasks are mapped” are units of work extracted from the application represented as IR traces and scheduled on processing elements. Performance evaluation of the different scheduling algorithms is performed using the disclosed simulation framework, thereby simulating a plurality of scheduling algorithms for a plurality of application tasks.);
during runtime on a heterogeneous system on a chip (SoC) wherein the policies are trained using the SoC simulator (Paragraph 125; “return the next state and rewards back to SOSPCS to help agents optimize themselves”, where the simulator provides the training feedback to train the policies. Instructions are executed in the simulator as applications run, and the simulators return the next state and reward signals at runtime, in which the feedback loop acts as labels for training the policies).
Morris, Higa, Matsuura, Li, Kroeger, and Bogdan are considered to be analogous to the claimed invention because they are in the same field of resource management using ML. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger to incorporate the teachings of Bogdan and employ a SoC simulator to simulate a plurality of scheduling algorithms for a plurality of tasks during runtime to train scheduler policies. A person of ordinary skill in the art would have recognized the advantages of simulation in reducing design costs, enabling evaluation of a plurality of scheduling strategies, and providing predictable improvements in task scheduling efficiency.
Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger further in view of Bogdan does not teach selecting, by an oracle, actions for task scheduling However, Khanna teaches:
selecting, by an oracle, actions for task scheduling (Paragraph 55; “system 102 may store feature vectors labeled through oracle identification… the labeled feature vectors can be used by a supervised machine learning system to train a machine learning model”, which is deployed to “facilitate classification and enforcement of policies” explicitly teaches an oracle labeling process used to predict actions, corresponding to policies, during runtime.);
Morris, Higa, Matsuura, Li, Kroeger, Bogdan, and Khanna are considered to be analogous to the claimed invention because they are in the same field of task management using ML. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Morris in view of Higa further in view of Bogdan to incorporate the teachings of Khanna and construct an oracle to train IL policies on. A person of ordinary skill in the art would understand that supervised training with oracle-provided labels is consistent with imitation learning practices in the art.
Regarding claim 12, Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger, further in view of Bogdan, further in view of Khanna teach the method of claim 11. Higa teaches:
IL (Paragraph 25; “The function that outputs an action to be taken by the agent according to the state of the target environment is called a policy. The imitation learning unit 30, described below, generates a policy through imitation learning.”).
Khanna teaches:
wherein the training of the policies comprises training the policies using supervised machine learning with the oracle such that the policies imitate the oracle (Paragraph 55; “the system 102 may store feature vectors labeled through oracle identification” where the vectors “can be used by a supervised machine learning system to train a machine learning model” explicitly discloses a supervised ML training process that is applied to the policies. “feature vectors labeled through oracle identification” explicitly teaches oracle identification being used to label data for training. “labeled feature vectors… used by a supervised machine learning system to train a machine learning model” corresponds to a supervised ML explicitly teaches that the learned policy/model imitates the oracle’s labels, because that is what it is trained on).
Bogdan teaches:
a heterogeneous SoC (Paragraphs 124-125; “a heterogeneous platform consisting of 32 CPUs and 32 GPUs is modeled using Macsim plus special HWAs with NoC substrate using BookSim2” discloses a simulator for heterogeneous SoC systems with CPUs, GPUs, and HWAs).
Morris teaches:
scheduling tasks at runtime (Paragraph 74-77; “The policy generator 410 may execute an actor network 411” and “may determine or select an action”. Further, “the policy-based scheduler 420 may perform first resource scheduling on the task 1 210 and the task a based on the determined action”, which discloses a policy generator producing policy network outputs, corresponding to imitation learning outputs, applied to scheduling tasks at runtime on a heterogeneous SoC. Paragraph 51 further discloses the heterogeneous SoC; “The resources 140 may include a plurality of clusters (for example, neural processing unit (NPU) clusters)”.);
generation of policies at runtime of the heterogeneous SoC (Paragraph 74; “The policy generator 410 may execute an actor network 411” which may “determine or select an action for the input vector” explicitly discloses a policy generator that produces policies for task scheduling. Further, “the policy-based scheduler 420 may perform first resource scheduling on the task 1 210 and the task a based on the determined action” teaches that the generated policies are applied to schedule tasks in real time on heterogeneous SoCs).
Regarding claim 16, Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger, further in view of Bogdan, further in view of Khanna teach the method of claim 11. Bogdan teaches:
wherein the SoC simulator is based on a heterogeneous SoC having heterogeneous processing elements grouped into different types of processing clusters (Paragraph 125; “Trace-driven timing simulators MacSim [23] and BookSim2 [16] execute these instructions and return the next state and rewards back to SOSPCS to help agents optimize themselves” explicitly disclosing a heterogeneous SoC simulator, the SOSPCS, that models workloads and scheduling. Bogdan further discloses “a heterogeneous platform consisting of 32 CPUs and 32 GPUs is modeled using MacSim plus special HWAs with NoC substrate using BookSim2” showing the heterogeneous SoC having a plurality of CPUs, GPUs, and accelerators, and is therefore heterogeneous. Further, “HWAs are specific in each domain, e.g., FFT and MM in signal processing or ReLU and neurons in NNs”. CPUs form one cluster, GPUs another, and HWAs are grouped into domain-specific clusters, each of which constitute different processing clusters within the heterogeneous SoC).
Regarding claim 19, Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger, further in view of Bogdan, further in view of Khanna teach the method of claim 16. Bogdan teaches:
wherein the heterogeneous SoC comprises one or more general processor clusters and one or more hardware accelerator clusters (Paragraph 125; “a heterogeneous platform consisting of 32 CPUs and 32 GPUs is modeled using MacSim plus special HWAs with NoC substrate using BookSim2. HWAs are specific in each domain, e.g., FFT and MM in signal processing or ReLU and neurons in NNs.” Explicitly discloses general purpose processors (CPUs) and hardware accelerators (everything else) in a heterogeneous SoC. The CPUs correspond to general processor clusters while the GPUs and domain-specific HWAs correspond to hardware accelerator clusters).
Regarding claim 20, Morris in view of Higa, further in view of Bogdan, further in view of Khanna, further in view of Nagpal teach the method of claim 19. Bogdan teaches:
wherein a hardware accelerator cluster of the one or more hardware accelerator clusters comprises at least one of: a cluster of matrix multipliers, a cluster of Viterbi decoders, a cluster of fast Fourier transform (FFT) accelerators, a cluster of graphical processing units (GPUs), a cluster of digital signal processors (DSPs), or a cluster of tensor processing units (TPUs) (Paragraph 125; “a heterogeneous platform consisting of 32 CPUs and 32 GPUs is modeled using MacSim plus special HWAs with NoC substrate using BookSim2. HWAs are specific in each domain, e.g., FFT and MM in signal processing or ReLU and neurons in NNs.” Covers the GPU element of the “at least one of” list of elements).
Claim 9 is rejected under 35 U.S.C. 103 as being unpatentable over Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger, further in view of Bogdan, further in view of Khanna, further in view of Pollack et al. (US 20150166265 A1) hereafter Pollack.
Regarding claim 9, Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger, further in view of Bogdan, further in view of Khanna teach the method of claim 5. Khanna teaches:
an oracle (Paragraph 55; “The labeled feature vectors can be used by a supervised machine learning system to train a machine learning model”. Since the labeled feature vectors originate from oracle identification, the training process therefore uses an oracle.).
Morris in view of Higa further in view of Bogdan further in view of Khanna does not teach being constructed from samples of multiple scheduling algorithms.
However, Pollack teaches:
samples of multiple scheduling algorithms (Paragraph 73; “In these embodiments, scheduling algorithms for the multiple modules should be coordinated to avoid conflicts for samples during a given operation cycle” explicitly shows multiple scheduling algorithms are present and used in coordination. The fact that they must be coordinated to handle samples implies the system uses their inputs, the samples of scheduling decisions, to guide scheduling).
Morris, Higa, Matsuura, Li, Kroeger, Bogdan, Khanna, and Pollack are considered to be analogous to the claimed invention because they are in the same field of information accessing for use in algorithms. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Morris in view of Higa further in view of Bogdan further in view of Khanna to incorporate the teachings of Pollack and have the oracle constructed from samples of multiple scheduling algorithms. A person of ordinary skill in the art would have recognized that combining multiple scheduler algorithms into a composite oracle produces a greater set of labels which would improve the robustness of the IL policies rather than relying on just one scheduler. By sampling multiple algorithms together, the system gains predictable fallback in edge cases.
Claims 13-15 are rejected under 35 U.S.C. 103 as being unpatentable over Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger, further in view of Bogdan, further in view of Khanna, further in view of Loginov et al. (US 11010776 B1) hereafter Loginov.
Regarding claim 13, Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger, further in view of Bogdan, further in view of Khanna teach the method of claim 12. Higa teaches:
IL (Paragraph 25; “The function that outputs an action to be taken by the agent according to the state of the target environment is called a policy. The imitation learning unit 30, described below, generates a policy through imitation learning.”).
Khanna teaches:
improving based on oracle actions (Paragraph 55; “feature vectors labeled through oracle identification or through an indictive learning process” which can then be “used by a supervised machine learning system to train a machine learning model” discloses oracle actions/labels being aggregated for ML training which improves the model).
Bogdan teaches:
aggregation of data from oracle actions and results of policies (Paragraph 125; “Trace driven timing simulators… execute these instructions and return the next state and rewards back to SOSPCS to help agents optimize themselves” explicitly shows results of IL policies being collected and used for improvement);
simulation (Paragraph 102; “During testing and simulation, unknown applications are used” to test the validity of the solution. The system explicitly utilizes simulation with machine learning applications).
Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger, further in view of Bogdan, further in view of Khanna does not teach using a data aggregator to improve policies based on actions during simulation.
However, Loginov teaches:
using a data aggregator to improve policies based on actions (Col. 7, lines 41-62; “services can be software modules running in working memory 214(1) and include a data aggregator 302… Policy/value reviser 308 utilizes machine learning and user feedback to review and revise object values and related data policy and to improve the valuation process” explicitly teaches a data aggregator that works with feedback to improve policies/values. The reviser that works with the aggregator that improves ML policies based on feedback corresponds to being utilized to improve policies).
Morris, Higa, Matsuura, Li, Kroeger, Bogdan, Khanna, and Loginov are considered to be analogous to the claimed invention because they are in the same field of resource management using ML. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger further in view of Bogdan further in view of Khanna to incorporate the teachings of Loginov and utilize a data aggregator to perform the actions of improving policies based on the results of actions taken during execution. A person of ordinary skill in the art would have recognized that this would have been a routine ML strategy for policy improvement resulting in the predictable outcome of improved policy accuracy and robustness.
Claim 14 recites similar limitations as the first limitation of claim 8, additionally reciting a data aggregator (DAgger). Loginov teaches:
a data aggregator (Col. 7, lines 41-62; “services can be software modules running in working memory 214(1) and include a data aggregator 302… Policy/value reviser 308 utilizes machine learning and user feedback to review and revise object values and related data policy and to improve the valuation process” explicitly teaches a data aggregator that works with feedback to improve policies/values).
Claim 14 is rejected for similar reasons as those of the first limitation of claim 8.
Regarding claim 15, Morris in view of Higa, further in view of Matsuura, further in view of Li, further in view of Kroeger, further in view of Bogdan, further in view of Khanna, further in view of Loginov teach the method of claim 13. Bogdan teaches:
improve the policies based on results of the IL policies during runtime on the heterogeneous SoC (Paragraphs 124-125; “Trace-driven timing simulators… execute these instructions and return the next state and rewards back to SOSPCS to help agents optimize themselves”, in which the simulators generate results in the form of state and reward signals which are used by the RL/IL agent to improve policies. “The experimental results are compared with state-of-the-art HETS scheduling”, and “The baseline for the results is to run applications with parallel execution using METIS in the simulator” where SOSPCS aggregates and compares oracle/baseline scheduling actions defined in HETS with the results. The benchmarking serves as oracle actions, providing labeled data for policy improvement. “execute these instructions and return the next state and rewards back to SOSPCS to help agents optimize” shows tasks executing with feedback at runtime. “heterogeneous platform consisting of 32 CPUs and GPUs” where simulation occurs explicitly discloses a heterogeneous SoC architecture).
Loginov teaches:
a data aggregator (Col. 7, lines 41-62; “services can be software modules running in working memory 214(1) and include a data aggregator 302… Policy/value reviser 308 utilizes machine learning and user feedback to review and revise object values and related data policy and to improve the valuation process” explicitly teaches a data aggregator that works with feedback to improve policies/values).
Response to Arguments
Applicant's arguments filed 01/23/2026 have been fully considered.
Applicant’s arguments are summarized below:
In view of the amendments to claim 8, reconsideration and withdrawal of the objections to claim 8 is requested.
In view of the amendments to the claims, withdrawal of the interpretation of the claims under 35 U.S.C. 112(f) and rejections under 112(a) and 112(b) is requested.
In view of the amendments to independent claims 1 and 11, the prior art of record does not teach, suggest, or render obvious the features of the independent claims.
Dependent claims are submitted as allowable for at least the above reasons.
In view of the amendments to the claims, the obviousness-type double patenting rejections of claims 1-3 and 5-6 should be withdrawn.
Examiner’s response:
The Examiner agrees that the amendments remedy the minor informalities of claim 8. Accordingly, the objection to claim 8 is withdrawn.
The Examiner agrees that the amendments to claims 11-15 avoid interpretation under 112(f). Accordingly, the interpretation of the phrases “IL policy generator” and “data aggregator (DAgger)” under 112(f) and associated rejections under 112(a) and 112(b) are withdrawn.
The Examiner agrees that the prior art of record does not teach, suggest, or render obvious the features of the independent claims in light of the amendments. Accordingly, the previous rejections under 35 U.S.C. 103 are withdrawn. However, upon further consideration, a new ground(s) of rejection is made in view of Morris, Higa, Matsuura, Li, and Kroeger, under 35 U.S.C. 103.
Independent claims 1 and 11 remain rejected for the reasons stated above. Therefore, contrary to Applicant's arguments, because the dependent claims depend from an unpatentable claim and does not add limitations that overcome the rejection, it likewise remains rejected.
Upon further review of the amendments to claim 1 in light of Radu (US 12332707 B2), the amended portions of claim 1 incorporate hierarchical IL policies to select processing elements and scheduling decisions based on specific scheduling parameters. In contrast, Radu is directed to hierarchical power management in a heterogeneous SoC including predicting power requirements using IL policies and adjusting execution time during runtime. The claims of Radu do not recite hierarchical cluster-based scheduling, multi-level processor selection policies, or scheduling based on queue ordering or task dependency constraints. Therefore, the amended claims are no longer directed to obvious variations of the claims of Radu. Accordingly, the obviousness-type double patenting rejection is withdrawn.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. Mitra et al. (US 11989647 B2) discloses a self-learning scheduler for resource requests utilizing reinforcement learning that iteratively learns a scheduling policy to improve scheduling distribution of resource requests on a compute infrastructure by learning characteristics and patterns of resource requests.
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 KENNETH P TRAN whose telephone number is (571)272-6926. The examiner can normally be reached M-TH 4:30 a.m. - 12:30 p.m. PT, F 4:30 a.m. - 8:30 a.m. PT, or at Kenneth.Tran@uspto.gov.
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, April Blair can be reached at (571) 270-1014. 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.
/KENNETH P TRAN/ Examiner, Art Unit 2196
/APRIL Y BLAIR/ Supervisory Patent Examiner, Art Unit 2196
Prosecution Insights