Granted Patent B2

US 12,437,253 · App. 18/147,802 · Granted Oct 7, 2025

Inventors: Sarath Sreedharan (Tempe, AZ); Tathagata Chakraborti (Cambridge, MA); Vinod Muthusamy (Austin, TX); Yara Rizk (Cambridge, MA); Yasaman Khazaeni (Needham, MA)

Assignee: International Business Machines Corporation

G06Q10/0633G06F9/4881

Abstract

Claims (81)

1. An automated workflow composer comprising-a processor system that performs processor system operations comprising:

using a workflow generator of the processor system to perform an automated workflow composition process that generates a composed workflow that includes logical and metric goal characteristics comprising the composed workflow, when executed by a host device in a domain, satisfying a target logical goal in a manner that optimizes a target unmodeled metric goal;

wherein the composed workflow further includes a workflow sequence structure comprising multiple tasks and multiple cases;

wherein the multiple tasks comprise atomic tasks, multi-instance tasks, and multi-instance block tasks;

wherein the multiple cases are configured to execute on the host device in parallel and independently of one another;

wherein the automated workflow composition process comprises performing iterations of a sequential workflow evaluation process;

wherein each iteration of the sequential workflow evaluation process advances the automated workflow process toward generating the composed workflow that includes the logical and metric goal characteristics comprising the composed workflow, when executed by the host device in the domain, satisfying the target logical goal in the manner that optimizes the target unmodeled metric goal;

wherein each iteration of the sequential workflow evaluation process comprises:

using the workflow generator and an optimization engine to perform workflow generator operations and optimization engine operations to generate a candidate workflow sequence structure that is a candidate for the composed workflow, the candidate workflow sequence structure having an associated candidate logical goal that is a candidate for satisfying the target logical goal;

using the workflow-metric model of the workflow generator to incorporate a workflow-metric function into the workflow sequence generation operations and the optimization engine operations, wherein the workflow-metric function generates a predicted metric that results from executing the candidate workflow sequence, wherein the predicted metric is a candidate for satisfying the target unmodeled metric goal; and

using the workflow metric model and the optimization engine to bias the workflow sequence generation operations toward the target unmodeled metric goal by using the predicted metric to determine how close the candidate workflow is to satisfying the target logical goal; and

wherein the workflow-metric model has been trained, using a historical workflow-metric corpus and a neural network that implements a sequential decision-making (SDM) algorithm and a long short-term memory (LSTM) algorithm, to uncover the workflow-metric function;

wherein the historical workflow-metric corpus is generated by:

accessing already-performed workflow sequences from multiple domains; and

labeling the already-performed workflow sequences with multiple types of metrics that measure performance results of the already-performed historical workflow sequences; and

wherein the target unmodeled metric goals are unmodeled in that the target unmodeled metric goals correspond to the multiple types of metrics that measure the performance results of the already-performed historical workflow sequences; and

responsive to generating the composed workflow, converting the composed workflow into host-executable instructions, storing the host-executable instructions in non-volatile memory of the host device for execution of parallel threads corresponding to the multiple cases, such that each thread executes its respective case independently and in parallel, and wherein execution of the composed workflow by the host device expressly modifies a memory state of the host device in real time as the workflow progresses.

2. The automated workflow composer of claim 1 , wherein the target unmodeled metric goal quantifies a performance feature of the composed workflow.

3. The automated workflow composer of claim 1 , wherein the workflow-metric function comprises learned relationships between workflow sequences and metric-based goals.

4. The automated workflow composer of claim 1 , wherein the workflow sequence generation operations comprise searching a plurality of combinations of possible workflows.

5. The automated workflow composer of claim 1 , wherein:

the host device comprises a conversational agent; and

the target logical goal comprises a state of the conversational agent that can be either true or false.

6. The automated workflow composer of claim 1 , wherein the automated workflow composition process generates the composed workflow based at least in part on the target logical goal, the target unmodeled metric goal, and features of the domain.

7. The automated workflow composer of claim 1 , wherein:

the target unmodeled metric goal comprises a key performance indicator (KPI) goal; and

the workflow-metric model comprises a sequence-to-value predictor.

8. A computer-implemented method comprising:

using a workflow generator comprising a processor system to perform an automated workflow composition process that generates a composed workflow that includes logical and metric goal characteristics comprising the composed workflow, when executed by a host device in a domain, satisfying a target logical goal in a manner that optimizes a target unmodeled metric goal;

wherein the composed workflow further includes a workflow sequence structure comprising multiple tasks and multiple cases;

wherein the multiple tasks comprise atomic tasks, multi-instance tasks, and multi-instance block tasks;

wherein the multiple cases are configured to execute on the host device in parallel and independently of one another; wherein the automated workflow composition process comprises performing iterations of a sequential workflow evaluation process;

wherein each iteration of the sequential workflow evaluation process advances the automated workflow process toward generating the composed workflow that includes the logical and metric goal characteristics comprising the composed workflow, when executed by the host device in the domain, satisfying the target logical goal in the manner that optimizes the target unmodeled metric goal;

wherein each iteration of the sequential workflow evaluation process comprises:

using the workflow generator and an optimization engine to perform workflow generator operations and optimization engine operations to generate a candidate workflow sequence structure that is a candidate for the composed workflow, the candidate workflow sequence structure having an associated candidate logical goal that is a candidate for satisfying the target logical goal;

using the workflow-metric model of the workflow generator to incorporate a workflow-metric function into the workflow sequence generation operations and the optimization engine operations, wherein the workflow-metric function generates a predicted metric that results from executing the candidate workflow sequence, wherein the predicted metric is a candidate for satisfying the target unmodeled metric goal; and

using the workflow metric model and the optimization engine to bias the workflow sequence generation operations toward the target unmodeled metric goal by using the predicted metric to determine how close the candidate workflow is to satisfying the target logical goal;

wherein the workflow-metric model has been trained, using a historical workflow-metric corpus and a neural network that implements a sequential decision-making (SDM) algorithm and a long short-term memory (LSTM) algorithm, to uncover the workflow-metric function;

wherein the historical workflow-metric corpus is generated by:

accessing already-performed workflow sequences from multiple domains; and

labeling the already-performed workflow sequences with multiple types of metrics that measure performance results of the already-performed historical workflow sequences; and

wherein the target unmodeled metric goals are unmodeled in that the target unmodeled metric goals correspond to the multiple types of metrics that measure the performance results of the already-performed historical workflow sequences; and

responsive to generating the composed workflow, converting the composed workflow into host-executable instructions, storing the host-executable instructions in non-volatile memory of the host device for execution of parallel threads corresponding to the multiple cases, such that each thread executes its respective case independently and in parallel, and wherein execution of the composed workflow by the host device expressly modifies a memory state of the host device in real time as the workflow progresses.

9. The computer-implemented method of claim 8 , wherein the target unmodeled metric goal quantifies a performance feature of the composed workflow.

10. The computer-implemented method of claim 8 , wherein the workflow-metric function comprises learned relationships between workflow sequences and metric-based goals.

11. The computer-implemented method of claim 8 , wherein the workflow sequence generation operations comprise searching a plurality of combinations of possible workflows.

12. The computer-implemented method of claim 8 , wherein:

the host device comprises a conversational agent; and

the target logical goal comprises a state of the conversational agent that can be either true or false.

13. The computer-implemented method of claim 8 , wherein the automated workflow composition process generates the composed workflow based at least in part on the target logical goal, the target unmodeled metric goal, and features of the domain.

14. The computer-implemented method of claim 8 , wherein:

the target unmodeled metric goal comprises a key performance indicator (KPI) goal; and

the workflow-metric model comprises a sequence-to-value predictor.

15. A computer program product comprising a computer readable program stored on a computer readable storage medium, wherein the computer readable program, when executed on a processor system, causes the processor system to perform processor system operations comprising:

using a workflow generator of the processor system to perform an automated workflow composition process that generates a composed workflow that includes logical and metric goal characteristics comprising the composed workflow, when executed by a host device in a domain, satisfying a target logical goal in a manner that optimizes a target unmodeled metric goal;

wherein the composed workflow further includes a workflow sequence structure comprising multiple tasks and multiple cases;

wherein the multiple tasks comprise atomic tasks, multi-instance tasks, and multi-instance block tasks;

wherein the multiple cases are configured to execute on the host device in parallel and independently of one another;

wherein the automated workflow composition process comprises performing iterations of a sequential workflow evaluation process;

wherein each iteration of the sequential workflow evaluation process advances the automated workflow process toward generating the composed workflow that includes the logical and metric goal characteristics comprising the composed workflow, when executed by the host device in the domain, satisfying the target logical goal in the manner that optimizes the target unmodeled metric goal;

wherein each iteration of the sequential workflow evaluation process comprises:

using the workflow generator and an optimization engine to perform workflow generator operations and optimization engine operations to generate a candidate workflow sequence structure that is a candidate for the composed workflow, the candidate workflow sequence structure having an associated candidate logical goal that is a candidate for satisfying the target logical goal;

using the workflow-metric model of the workflow generator to incorporate a workflow-metric function into the workflow sequence generation operations and the optimization engine operations, wherein the workflow-metric function generates a predicted metric that results from executing the candidate workflow sequence, wherein the predicted metric is a candidate for satisfying the target unmodeled metric goal; and

using the workflow metric model and the optimization engine to bias the workflow sequence generation operations toward the target unmodeled metric goal by using the predicted metric to determine how close the candidate workflow is to satisfying the target logical goal;

wherein the workflow-metric model has been trained, using a historical workflow-metric corpus and a neural network that implements a sequential decision-making (SDM) algorithm and a long short-term memory (LSTM) algorithm, to uncover the workflow-metric function;

wherein the historical workflow-metric corpus is generated by:

accessing already-performed workflow sequences from multiple domains; and

labeling the already-performed workflow sequences with multiple types of metrics that measure performance results of the already-performed historical workflow sequences; and

wherein the target unmodeled metric goals are unmodeled in that the target unmodeled metric goals correspond to the multiple types of metrics that measure the performance results of the already-performed historical workflow sequences; and

responsive to generating the composed workflow, converting the composed workflow into host-executable instructions, storing the host-executable instructions in non-volatile memory of the host device for execution of parallel threads corresponding to the multiple cases, such that each thread executes its respective case independently and in parallel, and wherein execution of the composed workflow by the host device expressly modifies a memory state of the host device in real time as the workflow progresses.

16. The computer program product of claim 15 , wherein the target unmodeled metric goal quantifies a performance feature of the composed workflow.

17. The computer program product of claim 15 , wherein:

the workflow-metric function comprises learned relationships between workflow sequences and metric-based goals; and

the workflow sequence generation operations comprise searching a plurality of combinations of possible workflows.

18. The computer program product of claim 15 , wherein:

the host device comprises a conversational agent; and

the target logical goal comprises a state of the conversational agent that can be either true or false.

19. The computer program product of claim 15 , wherein the automated workflow composition process generates the composed workflow based at least in part on the target logical goal, the target unmodeled metric goal, and features of the domain.

20. The computer program product of claim 15 , wherein:

the target unmodeled metric goal comprises a key performance indicator (KPI) goal; and

the workflow-metric model comprises a sequence-to-value predictor.

Assignments (1)

Continuity (1)

References Cited (81)