Method and system for optimizing delivery of consignments

A system (150) for optimizing delivery of consignments, the system (150) comprising:

• a memory (154) for storing instructions;
• and
• a processor (152) configured to execute the instructions and thereby cause the system (150) to at least:
• receive real-order data in relation to delivering a consignment comprising a plurality of packages, wherein the plurality of packages is to be delivered from at least one pickup location to a plurality of drop locations associated with a plurality of consignees and, wherein the real-order data comprises package related information corresponding to the plurality of packages and vehicle related information corresponding to a plurality of vehicles available for delivering the consignment,
• pre-process the package related information and the vehicle related information in the real-order data to generate a plurality of inputs,
• select a machine learning model (164) trained using deep reinforcement learning (DRL) to optimize an objective function of minimizing an overall cost of consignment delivery, wherein minimizing the overall cost comprises optimizing a number of vehicles selected for the consignment delivery based on an evaluation of a three-dimensional (3D) fitment of each package in loading bins associated with selected vehicles subject to a plurality of constraints associated with the consignment delivery and, wherein minimizing the overall cost further comprises optimizing a number of consignees and a number of drop locations serviced by each selected vehicle,
• provide the plurality of inputs to the machine learning model (164) comprising a plurality of input processing layers, to generate a combined feature space of all valid loading actions, wherein the plurality of input processing layers comprises a first feed-forward neural network (FFNN) layer, a second FFNN layer, a bi-directional long short-term memory (BiLSTM) layer, a first convolutional neural network (CNN) layer, and a second CNN layer, and wherein the plurality of input processing layers performs the following operations:
• generate, by the first FFNN layer, a first feature space (328) corresponding to a tensor embedding input representing information of available loading bin capacity of each selected vehicle;
• generate, by the BiLSTM layer, a second feature space (330) corresponding to a tensor embedding input representing information of remaining packages to be placed after a current time stamp, the BiLSTM layer configured to capture a sequential nature of package placements in the selected vehicles;
• generate, by the first CNN layer, a third feature space (332) corresponding to a tensor input representing a top view of a loading bin of each selected vehicle;
• and
• generate, by the second CNN layer, a fourth feature space (334) corresponding to a tensor input representing a top view of a current package to be placed at the current time stamp,
• wherein the machine learning model (164) is configured to predict a sequence of loading actions in relation to loading of the plurality of packages in the selected vehicles, wherein each loading action corresponds to placement of a selected package in a selected position in a loading bin of a selected vehicle from among the selected vehicles, and
• generate a loading plan based on the predicted sequence of loading actions, wherein the loading plan is configured to optimize the delivery of the plurality of packages associated with the consignment.

A computer-implemented method for optimizing delivery of consignments, the method comprising:

• receiving, by a system (150), real-order data in relation to delivering a consignment comprising a plurality of packages, wherein the plurality of packages is to be delivered from at least one pickup location to a plurality of drop locations associated with a plurality of consignees and, wherein the real-order data comprises package related information corresponding to the plurality of packages and vehicle related information corresponding to a plurality of vehicles available for delivering the consignment;
• preprocessing, by the system (150), the package related information and the vehicle related information in the real-order data to generate a plurality of inputs;
• selecting, by the system (150), a machine learning model (164) trained to use deep reinforcement learning (DRL) for optimizing packing of the plurality of packages in an optimum number of vehicles selected from among the plurality of vehicles, the packing of the plurality of packages in the selected vehicles performed subject to a plurality of constraints and subject to optimizing a number of consignees and a number of drop locations serviced by each selected vehicle;
• providing, by the system (150), the plurality of inputs to the machine learning model (164) comprising a plurality of input processing layers, to generate a combined feature space of all valid loading actions, wherein the plurality of input processing layers comprises a first feed-forward neural network (FFNN) layer, a second FFNN layer, a bi-directional long short-term memory (BiLSTM) layer, a first convolutional neural network (CNN) layer, and a second CNN layer, and wherein the plurality of input processing layers performs the following operations:
• generating, by the first FFNN layer, a first feature space (328) corresponding to a tensor embedding input representing information of available loading bin capacity of each selected vehicle;
• generating, by the BiLSTM layer, a second feature space (330) corresponding to a tensor embedding input representing information of remaining packages to be placed after a current time stamp, the BiLSTM layer configured to capture a sequential nature of package placements in the selected vehicles;
• generating, by the first CNN layer, a third feature space (332) corresponding to a tensor input representing a top view of a loading bin of each selected vehicle;
• and
• generating, by the second CNN layer, a fourth feature space (334) corresponding to a tensor input representing a top view of a current package to be placed at the current time stamp,
• wherein the machine learning model (164) is configured to predict a sequence of loading actions in relation to loading of the plurality of packages in the selected vehicles, wherein each loading action corresponds to placement of a selected package in a selected position in a loading bin of a selected vehicle from among the selected vehicles;
• and
• generating, by the system (150), a loading plan based on the predicted sequence of loading actions, wherein the loading plan is configured to optimize the delivery of the plurality of packages associated with the consignment.