Ride Dispatch Neural Network for Dynamic Vehicle Repositioning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current vehicle dispatch platforms face challenges in efficiently allocating drivers to ride requests and repositioning them due to the exploration-exploitation dilemma and delayed consequences of assignment actions, leading to suboptimal short-term and long-term effects on revenue and driver availability.
Innovation Solution
The implementation of a single-driver deep-Q network (SD-DQN) and multi-driver deep-Q network (MD-DQN) architecture that uses a global state representation to learn action-value functions for non-myopic dispatching and repositioning, optimizing driver assignments and repositioning decisions based on Q-values calculated through neural networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional vehicle dispatch platforms use simple allocation algorithms, then the system complexity is low and ease of operation is maintained, but the productivity and revenue optimization is insufficient due to the exploration-exploitation dilemma
Solution Approach 1:
The patent replaces traditional mechanical dispatch algorithms with a neural network-based deep Q-learning system. The neural network learns optimal dispatch policies through reinforcement learning, substituting conventional rule-based or heuristic allocation methods with an intelligent agent that adapts to dynamic conditions, thereby resolving the exploration-exploitation dilemma and improving revenue optimization.
Solution Approach 2:
The system transforms the dispatch problem into a parameter optimization problem by using Q-values to evaluate different dispatch actions. The neural network learns to optimize parameters such as driver selection, repositioning decisions, and allocation strategies by maximizing cumulative rewards, enabling the system to balance exploration and exploitation dynamically.
2Ease of operation
If the dispatch platform makes decisions based on immediate short-term impact, then the ease of operation is maintained, but the long-term effects on driver distribution and future order service capability deteriorate
Solution Approach 1:
The deep Q-learning system performs preliminary learning of optimal dispatch policies through extensive training before actual operation. During training, the system explores various scenarios and learns long-term strategies for driver allocation and repositioning. This preliminary action enables the system to make decisions that balance immediate needs with long-term driver distribution, improving reliability for future order service.
Solution Approach 2:
The system implements feedback mechanisms where the neural network receives rewards or penalties based on the long-term outcomes of dispatch decisions. This feedback loop allows the system to learn from both short-term results and long-term driver availability patterns, adjusting its policy to maintain driver distribution that supports future order service capability while maintaining operational simplicity.
3Productivity
If the platform uses complex deep-Q network models for dispatching, then the productivity and revenue are optimized, but the device complexity and computational requirements increase
Solution Approach 1:
The patent segments the complex multi-driver dispatch problem into manageable components by using a centralized neural network that processes individual driver states and actions. The deep Q-learning system evaluates Q-values for each driver independently based on global state information, breaking down the complex optimization problem into smaller sub-problems that can be solved efficiently while maintaining overall system optimization.
4Device complexity
If traditional dispatch methods are used, then the device complexity is low, but the loss of time occurs due to delayed consequences of assignment actions not being properly accounted for
Solution Approach 1:
The patent replaces traditional sequential or rule-based dispatch algorithms with a neural network-based system that can evaluate multiple potential actions and their long-term consequences simultaneously. The deep Q-learning model uses learned value functions to predict future outcomes, enabling faster and more accurate dispatch decisions that account for delayed consequences without requiring complex iterative calculations.
Data Source
AI summary
Systems and methods are provided for ride order dispatching and vehicle repositioning. A method for ride order dispatching and vehicle repositioning, comprises: obtaining information comprising a location of a vehicle, current orders, and a current time; inputting the obtained information to a trained model; and determining action information for the vehicle based on an output of the trained model, the action information comprising: re-positioning the vehicle or accepting a ride order. The model is configured with: receiving information of drivers and information of orders as inputs; obtaining a global state based on the information of drivers, the information of orders, and a global time; and querying a plurality of driver-order pairs and driver-reposition pairs based at least on the obtained global state to determine the action information as the output.