Multi-agent reinforcement learning for order-dispatching via order-vehicle distribution matching
View Patent ↗Multi-agent reinforcement learning may be used for rider order-dispatching via matching the distribution of orders and vehicles. Information may be obtained. The information may include a plurality of vehicle locations of a plurality of vehicles, a plurality of ride orders, and a current time. The obtained information may be input into a trained model. The trained model may be based on Kullback-Leibler divergence optimization and independent agents under a guidance of a joint policy. A plurality of order-dispatching tasks may be generated for the plurality of vehicles to fulfill.
1. A method for ride order-dispatching, comprising:
training a Q-learning network that minimizes vehicle waiting time and traveling time in a ride-sharing platform, wherein the training comprises using Kullback-Leibler divergence optimization and independent agents under a guidance of a joint policy, and the Q-learning network represents the joint policy;
obtaining information comprising a plurality of vehicle locations of a plurality of vehicles, pending ride orders, and a current time;
inputting the obtained information as a state into the trained Q-learning network to obtain recommended vehicle actions, wherein the vehicle actions comprise accepting an order or staying idle, wherein:
the trained Q-learning network adds a virtual order when a number of the pending orders is zero, and
when a vehicle's recommended vehicle action include accepting the virtual order, the vehicle stays idle; and
generating a plurality of order-dispatching tasks for the plurality of vehicles to fulfill.
2. The method of claim 1 , wherein the method further comprises:
obtaining a grid-world representing a real world geographical area, wherein the grid-world comprises a plurality of grid cells representing a plurality of order-dispatching regions, wherein each vehicle location of the plurality of vehicle locations comprises a grid in the grid-world.
3. The method of claim 2 , wherein the method further comprises:
obtaining a plurality of states for the plurality of vehicles, wherein each state comprises a grid cell in the grid-world, a number of idle vehicles in the grid cell, a number of ride orders in the grid cell, and a distribution of ride orders in the grid cell.
4. The method of claim 3 , wherein the plurality of order-dispatching tasks are generated by selecting ride orders from the plurality of ride orders in the plurality of states.
5. The method of claim 4 , wherein each selected ride order comprises the grid cell from the corresponding state, a target grid cell, a ride duration, and a price.
6. The method of claim 1 , wherein the Q learning network uses biased strategy Boltzmann exploration to balance exploitation and exploration.
7. The method of claim 1 , wherein the Kullback-Leibler divergence optimization comprises minimizing a divergence between a distribution of vehicles and a distribution of ride orders.
8. The method of claim 1 , wherein the Q learning network is trained using centralized learning.
9. A system for ride order-dispatching, comprising one or more processors and one or more non-transitory computer-readable memories coupled to the one or more processors and configured with instructions executable by the one or more processors to cause the system to perform operations comprising:
training a Q-learning network that minimizes vehicle waiting time and traveling time in a ride-sharing platform, wherein the training comprises using Kullback-Leibler divergence optimization and independent agents under a guidance of a joint policy, and the Q-learning network represents the joint policy;
obtaining information comprising a plurality of vehicle locations of a plurality of vehicles, pending ride orders, and a current time;
inputting the obtained information as a state into the trained Q-learning network to obtain recommended vehicle actions, wherein the vehicle actions comprise accepting an order or staying idle, wherein:
the trained Q-learning network adds a virtual order when a number of the pending orders is zero, and
when a vehicle's recommended vehicle action include accepting the virtual order, the vehicle stays idle; and
generating a plurality of order-dispatching tasks for the plurality of vehicles to fulfill.
10. The system of claim 9 , wherein the operations further comprise:
obtaining a grid-world representing a real world geographical area, wherein the grid-world comprises a plurality of grid cells representing a plurality of order-dispatching regions, wherein each vehicle location of the plurality of vehicle locations comprises a grid in the grid-world.
11. The system of claim 10 , wherein the operations further comprise:
obtaining a plurality of states for the plurality of vehicles, wherein each state comprises a grid in the grid-world, a number of idle vehicles in the grid, a number of ride orders in the plurality of ride orders, and a distribution of ride orders in the grid cell.
12. The system of claim 11 , wherein the plurality of order-dispatching tasks are generated by selecting ride orders from the plurality of ride orders in the plurality of states.
13. The system of claim 12 , wherein each selected ride order comprises the grid cell from the corresponding state, a target grid cell, a ride duration, and a price.
14. The system of claim 9 , wherein the Q-learning network uses biased strategy Boltzmann exploration to balance exploitation and exploration.
15. The system of claim 9 , wherein the Kullback-Leibler divergence optimization comprises minimizing a divergence between a distribution of vehicles and a distribution of ride orders.
16. A non-transitory computer-readable storage medium configured with instructions executable by one or more processors to cause the one or more processors to perform operations comprising:
training a Q-learning network that minimizes vehicle waiting time and traveling time in a ride-sharing platform, wherein the training comprises using Kullback-Leibler divergence optimization and independent agents under a guidance of a joint policy, and the Q-learning network represents the joint policy;
obtaining information comprising a plurality of vehicle locations of a plurality of vehicles, pending ride orders, and a current time;
inputting the obtained information as a state into the trained Q-learning network to obtain recommended vehicle actions, wherein the vehicle actions comprise accepting an order or staying idle, wherein:
the trained Q-learning network adds a virtual order when a number of the pending orders is zero, and
when a vehicle's recommended vehicle action include accepting the virtual order, the vehicle stays idle; and
generating a plurality of order-dispatching tasks for the plurality of vehicles to fulfill.
17. The non-transitory computer-readable storage medium of claim 16 , wherein the operations further comprise:
obtaining a grid-world representing a real world geographical area, wherein the grid-world comprises a plurality of grid cells representing a plurality of order-dispatching regions, wherein each vehicle location of the plurality of vehicle locations comprises a grid in the grid-world.
18. The non-transitory computer-readable storage medium of claim 17 , wherein the operations further comprise:
obtaining a plurality of states for the plurality of vehicles, wherein each state comprises a grid in the grid-world, a number of idle vehicles in the grid, a number of ride orders in the plurality of ride orders, and a distribution of ride orders in the grid cell.
19. The non-transitory computer-readable storage medium of claim 16 , wherein the Kullback-Leibler divergence optimization comprises minimizing a divergence between a distribution of vehicles and a distribution of ride orders.