Ally-adversary bimodal resource allocation optimization

1 . A method, in a data processing system, for generating a resource allocation recommendation in an omnichannel distribution network having a plurality of locations and channels for moving resources between locations, the method comprising:

retrieving historical and current inventory data related to a target resource and the omnichannel distribution network;

generating, using a machine learning model, a hierarchical demand forecast data, for allocation of resources within the omnichannel distribution network, based at least in part on the historical and current inventory data, wherein the hierarchical demand forecast data comprises online demand computed over a zone comprising a geographic cluster;

instantiating an ally-adversary bimodal inventory optimization (BIO) computer model comprising an adversary component that simulates, through a first computer simulation, a worst-case scenario of resource demand and resource availability, and an ally component that limits the adversary component based on a second computer simulation of a limited best-case scenario of resource demand and resource availability, wherein the BIO computer model is instantiated within a virtual computing environment comprising a plurality of virtual machines and container sets;

executing, by the data processing system, the first computer simulation and the second computer simulation of the BIO computer model based on the hierarchical demand forecast data and current inventory data as input to the BIO computer model, to generate a predicted consumption for the target resource across a plurality of locations in the omnichannel distribution network, wherein the plurality of locations within the omnichannel distribution network comprises one or more of warehouses, retail stores, or electronic commerce fulfillment centers, and wherein executing the first computer simulation and the second computer simulation comprises performing, by the data processing system, one or more iterations associated with an objective function, wherein each iteration comprises:

updating a first bound of the objective function based on the limited best-case scenario; and

updating a second bound of the objective function based on the worst-case scenario;

generating the resource allocation recommendation for allocating the resource to locations of the omnichannel distribution network, based on the predicted consumption for the target resource;

outputting the resource allocation recommendation to a downstream computing system to perform a downstream computing operation based on the resource allocation recommendation, wherein the downstream computing system is an enterprise resource planning (ERP) system for inventory management; and

performing a redistribution of physical inventory at one or more nodes of the omnichannel distribution network based at least in part on the resource allocation recommendation.

2 . The method of claim 1 ,

wherein the worst-case scenario simulates demand being misaligned with resource inventory in the omnichannel distribution network based on a worst-case demand uncertainty and fulfillment of a predicted worst-case demand, and wherein the limited best-case scenario simulates one or more portions of the omnichannel distribution network having demand aligned with resource inventory.

3 . The method of claim 2 ,

wherein the BIO computer model is configured with a first hyperparameter K that limits a number of portions of the omnichannel distribution network that has demand aligned with resource inventory, and a second hyperparameter λ that is a fraction of a demand budget constraint representing a best-case demand portion of a demand budget.

4 . The method of claim 1 ,

wherein the adversary component searches over all possible demand realizations in an uncertainty set controlled by a budget constraint, and wherein the ally component, for at least a selected subset of locations within the omnichannel distribution network, simulates a resource supply at that selected subset of locations meeting resource demand at the selected subset of locations.

5 . The method of claim 1 ,

wherein the ally component limits the worst-case scenario simulated by the adversary component based on a demand fraction of a budget constraint.

6 . The method of claim 1 ,

wherein the BIO computer model simulates an allocation decision problem having three stages comprising:

in stage 1, an allocation of resources is made based on an allied demand determination that aligns resource inventory to demand at locations within the omnichannel distribution network;

in stage 2, the adversary component selects a worst-case adversarial demand condition based on the allocation from stage 1; and

in stage 3, the allocation of resources is modified to fulfill demands based on the demands modeled in stages 1 and 2.

7 . The method of claim 6 ,

wherein stage 1 is processed as a master problem using a Benders Decomposition technique to model worst-case scenarios, and wherein stages 2 and 3 are combined to generate a complex stage 2 that is processed as a subproblem using the Benders Decomposition technique to generate scenarios to feed to the master problem, and wherein the master problem and subproblem are processed iteratively for different scenarios generated by the subproblem until the master problem and subproblem converge.

8 . The method of claim 1 ,

wherein applying the BIO computer model to the hierarchical demand forecast data and current inventory data, comprises:

solving a master problem to calculate tentative inventory purchase quantity and allocations;

updating a best-case demand to all locations within the omnichannel distribution network, and an upper bound value;

solving a sub-problem to identify worst-case demands at all locations within the omnichannel distribution network, and a lower bound value; and

iterating the solving of the master problem, the updating, and solving the sub-problem until the upper bound value and the lower bound value converge.

9 . The method of claim 8 , wherein, in response to the upper bound value and lower bound value converging, scoring resource purchase and allocations of buy quantity based on a validation data set.

10 . The method of claim 1 ,

wherein the target resource is a good or product.

11 . A non-transitory computer-readable medium storing a set of instructions for generating a resource allocation recommendation in an omnichannel distribution network, the set of instructions comprising:

one or more instructions that, when executed by one or more processors of a device, cause the device to:

retrieve historical and current inventory data related to a target resource and the omnichannel distribution network;

generate, using a machine learning model, a hierarchical demand forecast data, for allocation of resources within the omnichannel distribution network, based at least in part on the historical and current inventory data, wherein the hierarchical demand forecast data comprises online demand computed over a zone comprising a geographic cluster;

instantiate an ally-adversary bimodal inventory optimization (BIO) computer model comprising an adversary component that simulates, through a first computer simulation, a worst-case scenario of resource demand and resource availability, and an ally component that limits the adversary component based on a second computer simulation of a limited best-case scenario of resource demand and resource availability, wherein the BIO computer model is instantiated within a virtual computing environment comprising a plurality of virtual machines and container sets;

execute the first computer simulation and the second computer simulation of the BIO computer model based on the hierarchical demand forecast data and current inventory data as input to the BIO computer model, to generate a predicted consumption for the target resource across a plurality of locations in the omnichannel distribution network, wherein the plurality of locations within the omnichannel distribution network comprises one or more of warehouses, retail stores, or electronic commerce fulfillment centers, and wherein, to execute the first computer simulation and the second computer simulation, the one or more instructions cause the device to perform one or more iterations associated with an objective function, wherein each iteration comprises:

updating a first bound of the objective function based on the limited best-case scenario; and

updating a second bound of the objective function based on the worst-case scenario;

generate the resource allocation recommendation for allocating the target resource to locations of the omnichannel distribution network, based on the predicted consumption for the target resource;

output the resource allocation recommendation to a downstream computing system to perform a downstream computing operation based on the resource allocation recommendation, wherein the downstream computing system is an enterprise resource planning (ERP) system for inventory management; and

perform a redistribution of physical inventory at one or more nodes of the omnichannel distribution network based at least in part on the resource allocation recommendation.

12 . The non-transitory computer-readable medium of claim 11 ,

wherein the worst-case scenario simulates demand being misaligned with resource inventory in the omnichannel distribution network based on a worst-case demand uncertainty and fulfillment of a predicted worst-case demand, and wherein the limited best-case scenario simulates one or more portions of the omnichannel distribution network having demand aligned with resource inventory.

13 . The non-transitory computer-readable medium of claim 12 ,

wherein the one or more instructions cause the device to configure the BIO computer model with a first hyperparameter K that limits a number of portions of the omnichannel distribution network that has demand aligned with resource inventory, and a second hyperparameter λ that is a fraction of a demand budget constraint representing a best-case demand portion of a demand budget.

14 . The non-transitory computer-readable medium of claim 11 ,

wherein the one or more instructions cause the adversary component to search over all possible demand realizations in an uncertainty set controlled by a budget constraint, and wherein one or more instructions cause the ally component to simulate, for at least a selected subset of locations within the omnichannel distribution network, a resource supply at that selected subset of locations meeting resource demand at the selected subset of locations.

15 . The non-transitory computer-readable medium of claim 11 ,

wherein the ally component limits the worst-case scenario simulated by the adversary component based on a demand fraction of a budget constraint.

16 . The non-transitory computer-readable medium of claim 11 ,

wherein the one or more instructions cause the BIO computer model to simulate an allocation decision problem having three stages comprising:

in stage 1, an allocation of resources is made based on an allied demand determination that aligns resource inventory to demand at locations within the omnichannel distribution network as much as possible;

in stage 2, the adversary component selects a worst-case adversarial demand condition based on the allocation from stage 1; and

in stage 3, the allocation of resources is modified to fulfill demands based on the demands modeled in stages 1 and 2.

17 . The non-transitory computer-readable medium of claim 16 ,

wherein stage 1 is processed as a master problem using a Benders Decomposition technique to model worst-case scenarios, and wherein stages 2 and 3 are combined to generate a complex stage 2 that is processed as a subproblem using the Benders Decomposition technique to generate scenarios to feed to the master problem, and wherein the master problem and subproblem are processed iteratively for different scenarios generated by the subproblem until the master problem and subproblem converge.

18 . The non-transitory computer-readable medium of claim 11 ,

wherein to apply the BIO computer model to the hierarchical demand forecast data and current inventory data, the one or more instructions cause the device to:

solve a master problem to calculate tentative inventory purchase quantity and allocations;

update a best-case demand to all locations within the omnichannel distribution network, and an upper bound value;

solve a sub-problem to identify worst-case demands at all locations within the omnichannel distribution network, and a lower bound value; and

iterate the solving of the master problem, the updating, and solving the sub-problem until the upper bound value and the lower bound value converge.

19 . The non-transitory computer-readable medium of claim 18 , wherein, in response to the upper bound value and lower bound value converging, the one or more instructions cause the device to score resource purchase and allocations of buy quantity based on a validation data set.

20 . An apparatus comprising:

one or more processors; and

one or more memory devices coupled to the one or more processors, wherein the one or more processors are configured to:

retrieve historical and current inventory data related to a target resource and an omnichannel distribution network;

generate, with a machine learning model, a hierarchical demand forecast data, for allocation of resources within the omnichannel distribution network, based at least in part on the historical and current inventory data, wherein the hierarchical demand forecast data comprises online demand computed over a zone comprising a geographic cluster;

instantiate an ally-adversary bimodal inventory optimization (BIO) computer model comprising an adversary component that simulates, through a first computer simulation, a worst-case scenario of resource demand and resource availability, and an ally component that limits the adversary component based on a second computer simulation of a limited best-case scenario of resource demand and resource availability, wherein the BIO computer model is instantiated within a virtual computing environment comprising a plurality of virtual machines and container sets;

execute, by the one or more processors, the first computer simulation and the second computer simulation of the BIO computer model based on the hierarchical demand forecast data and current inventory data as input to the BIO computer model, to generate a predicted consumption for the target resource across a plurality of locations in the omnichannel distribution network, wherein the plurality of locations within the omnichannel distribution network comprises one or more of warehouses, retail stores, or electronic commerce fulfillment centers, and wherein, to execute the first computer simulation and the second computer simulation, the at least on processor is configured to perform one or more iterations associated with an objective function, wherein each iteration comprises:

updating a first bound of the objective function based on the limited best-case scenario; and

updating a second bound of the objective function based on the worst-case scenario;

generate a resource allocation recommendation for allocating the target resource to locations of the omnichannel distribution network, based on the predicted consumption for the target resource;

output the resource allocation recommendation to a downstream computing system to perform a downstream computing operation based on the resource allocation recommendation, wherein the downstream computing system is an enterprise resource planning (ERP) system for inventory management; and

perform a redistribution of physical inventory at one or more nodes of the omnichannel distribution network based at least in part on the resource allocation recommendation.