CN107919014B

CN107919014B - Taxi running route optimization method for multiple passenger mileage

Info

Publication number: CN107919014B
Application number: CN201711103910.2A
Authority: CN
Inventors: 荣辉桂; 张群; 杨昌; 张旭东; 潘梦颖
Original assignee: Hunan University
Current assignee: Hunan University
Priority date: 2017-11-10
Filing date: 2017-11-10
Publication date: 2020-10-30
Anticipated expiration: 2037-11-10
Also published as: CN107919014A

Abstract

The invention discloses a taxi income efficiency optimization method facing multi-passenger mileage, which comprises the steps of dividing a taxi operation area into a plurality of grids, and simultaneously acquiring historical operation data in the operation area; positioning the current position and operation state of a taxi; and optimizing the next operation route of the taxi, thereby finishing the income efficiency optimization of the taxi hiring. The method takes the maximum income of a single taxi as a target, comprehensively considers factors such as historical data, statistical probability, cost and income, and dynamically optimizes the driving path of the taxi by adopting a Markov decision process method, so that the income maximization of the single taxi is ensured.

Description

Taxi running route optimization method for multiple passenger mileage

Technical Field

The invention particularly relates to a taxi running route optimization method for multiple passenger mileage.

Background

With the development of national economic technology and the improvement of living standard of people, the taxi industry is also greatly developed. Taxi service plays an important role in urban public transport systems and is an indispensable important part in urban public transport.

However, a phenomenon common to the taxi industry is: a small part of taxis are in a state of being used by passengers, a small part of taxi drivers are in a passenger state of parking in a large junction area (such as a railway station, a bus station and the like), and a large part of taxi drivers blindly drive on the road so as to meet the passengers needing to rent the taxis. Therefore, in the existing taxi industry, the common problem is that the passenger source cannot be found quickly, so that the empty driving rate of the taxi is reduced, the utilization rate of the taxi is improved, the income of a taxi driver is improved, and the invalid energy consumption of the taxi is reduced.

At present, although some routes of taxis are optimized, the routes are mainly recommended for the next mileage search, and the service efficiency of the taxis is improved; the existing research is not comprehensive enough in consideration, a method for continuously maximizing the income of the taxi in a plurality of mileage within a period of time is lacked, and the practical application value is not high.

Disclosure of Invention

The invention aims to provide a taxi running route optimization method which is relatively comprehensive in consideration and designed for individual taxis and is oriented to multiple passenger mileage.

The invention provides a taxi running route optimization method for multiple passenger mileage, which comprises the following steps:

s1, dividing an operation area of a taxi into a plurality of grids, and simultaneously acquiring historical operation data in the operation area;

s2, positioning the current position and operation state of the taxi;

and S3, optimizing the next operation route of the taxi according to the current position and the operation state of the taxi acquired in the step S2, thereby finishing the income efficiency optimization of the taxi.

Step S1, performing meshing on the operation area of the taxi, specifically performing meshing by using the following principle: the size of the divided grid area can be matched with road section information in the operation area of the taxi.

The historical operation data in the operation area in step S1 includes the average time from grid i to grid j at time t, the average income for transporting a passenger from grid i to grid j at time t, the time for traveling through grid i at time t, the number of taxis of the passenger found in grid i at time t, the number of taxis passing through grid i at time t, the number of passengers using area j as a destination at time t, and the like, where t is a positive real number, and i and j are both natural numbers.

The step S3 of optimizing the next operation route of the taxi specifically includes the following steps:

A. obtaining the current operation state of a taxi;

B. and D, optimizing the next operation route of the taxi by adopting a Markov decision process method according to the operation state of the taxi acquired in the step A.

And B, optimizing the next operation route of the taxi by adopting the following rules:

r1, if the taxi is in a state of transporting passengers, optimizing a next-step operation route of the taxi by adopting the existing mature city path optimization algorithm;

r2, if the taxi is in the non-passenger-carrying state and the duration time of the non-passenger-carrying state is greater than or equal to the set threshold value, determining that the taxi is in the non-operation state, and stopping optimizing the next operation route of the taxi;

and R3, if the taxi is in the passenger-free state and the duration time of the passenger-free state is less than a set threshold value, determining that the taxi is in a passenger searching state, and optimizing the next operation route of the taxi by taking the following function as a target function:

in the formula, max is a function of solving the maximum value, and V (s, a) is expected yield which can be brought by action a of the taxi in a state s in the current time period; p_find(l) Probability of a taxi being able to find a passenger in grid l, V (l)_a,t+t_seek(l_a) 10-a) is a taxi and has spent t + t_seek(l_a) Time and in the current grid area l_aIf no passenger is found, taking (10-a) as the recommended expected value of the driving direction; v (k, t + t)_seek(j)+t_drive(j, k),0) represents the expected value of finding a passenger in the j region, where t + t_seek(j)+t_drive(j, k) represents finding a passenger in the j area to go to the k area, when the time spent searching and the time spent from the j road section to the k road section to be spent are about to be spent; t is the current time, t_seek(l_a) For touring target grid area l_aRequired time, P_dest(j, k) is the probability from grid area j to grid area k, r (j, k) is taxi from gridExpected benefit, t, of region j to grid region k_drive(j, k) is the time it takes for the taxi to go from grid area j to grid area k; a is an action taken by the vacated taxi, a belongs to A, and A is an action set which can be taken by the vacated taxi; a is a natural number of 1-9, wherein 1 represents that the next action of the taxi is driving from the current position to the southwest direction, 2 represents that the next action of the taxi is driving from the current position to the southwest direction, 3 represents that the next action of the taxi is driving from the current position to the southeast direction, 4 represents that the next action of the taxi is driving from the current position to the southwest direction, 5 represents that the next action of the taxi is staying at the current position, 6 represents that the next action of the taxi is driving from the current position to the southwest direction, 7 represents that the next action of the taxi is driving from the current position to the northwest direction, 8 represents that the next action of the taxi is driving from the current position to the northwest direction, and 9 represents that the next action of the taxi is driving from the current position to the northeast direction.

The optimization of the next operation route of the taxi according to the rule R3 is specifically to solve the current function by using a dynamic planning method, so as to optimize the next operation route of the taxi.

The method for optimizing the taxi running route facing the multi-passenger mileage comprehensively considers factors such as historical data, statistical probability, cost and income by taking the maximum income of a plurality of miles of a single taxi in a period of time as a target, and dynamically optimizes the running path of the taxi by adopting a Markov decision process method, so that the income maximization of the single taxi is ensured.

Drawings

FIG. 1 is a process flow diagram of the process of the present invention.

Fig. 2 is a schematic view of a hot spot of a taxi getting-on position at different time intervals in the embodiment of the method.

FIG. 3 is a bar graph comparing top 10% ranked drivers to bottom 10% ranked drivers in an embodiment of the method of the present invention.

Fig. 4 is a schematic diagram of a set of actions that can be taken by a taxi in accordance with the method of the present invention.

FIG. 5 is a diagram illustrating simulation comparison of the method of the present invention with historical data.

FIG. 6 is a diagram illustrating a simulation comparison of the method of the present invention with historical data during daytime.

Detailed Description

FIG. 1 shows a flow chart of the method of the present invention: the invention provides a taxi running route optimization method for multiple passenger mileage, which comprises the following steps:

s1, dividing an operation area of a taxi into a plurality of grids, and simultaneously obtaining historical operation data in the operation area, wherein the historical operation data comprises average time from grid i to grid j at time t, average income of passengers transported from grid i to grid j at time t, time for patrolling grid i at time t, the number of taxis of the passengers found in grid i at time t, the number of taxis passing through grid i at time t, the number of passengers taking area j as a destination at time t and the like, wherein t is a positive real number, and i and j are natural numbers;

s2, positioning the current position and operation state of the taxi;

s3, optimizing the next operation route of the taxi according to the current position and the operation state of the taxi acquired in the step S2, so that the income efficiency optimization of the taxi is completed; specifically, the following steps are adopted to optimize the next operation route of the taxi:

A. obtaining the current operation state of a taxi;

B. b, optimizing the next operation route of the taxi by adopting a Markov decision process method according to the operation state of the taxi acquired in the step A; specifically, the following rules are adopted for optimization:

in the formula, max is a function of solving the maximum value, and V (s, a) is expected yield which can be brought by action a of the taxi in a state s in the current time period; p_find(l) Probability of a taxi being able to find a passenger in grid l, V (l)_a,t+t_seek(l_a) 10-a) T + t has been spent for the taxi_seek(l_a) Time and in the current grid area l_aIf no passenger is found, taking (10-a) as the recommended expected value of the driving direction; v (k, t + t)_seek(j)+t_drive(j, k),0) represents the expected value of finding a passenger in the j region, where t + t_seek(j)+t_drive(j, k) represents finding a passenger in the j area to go to the k area, when the time spent searching and the time spent from the j road section to the k road section to be spent are about to be spent; t is the current time, t_seek(l_a) For touring target grid area l_aRequired time, P_dest(j, k) is the probability from grid area j to grid area k, r (j, k) is the expected revenue for taxis from grid area j to grid area k, t_drive(j, k) is the time it takes for the taxi to go from grid area j to grid area k; a is an action taken by the vacated taxi, a belongs to A, and A is an action set which can be taken by the vacated taxi; a is a natural number of 1-9, 1 represents that the next action of the taxi is driving from the current position to the southwest direction, 2 represents that the next action of the taxi is driving from the current position to the southwest direction, 3 represents that the next action of the taxi is driving from the current position to the southeast direction, and 4 represents that the taxi is rentedThe next action of the taxi is to drive from the current position to the northwest direction, 5 indicates that the next action of the taxi is to stay at the current position, 6 indicates that the next action of the taxi is to drive from the current position to the northwest direction, 7 indicates that the next action of the taxi is to drive from the current position to the northwest direction, 8 indicates that the next action of the taxi is to drive from the current position to the northwest direction, and 9 indicates that the next action of the taxi is to drive from the current position to the northeast direction.

Meanwhile, a dynamic planning method can be adopted to solve the current function, so that the next operation route of the taxi is optimized.

The process of the invention is further illustrated below with reference to a specific example:

the taxi position data comprises taxi operation information of the whole H city area, the research area is set as an urban area range, and data exceeding the range is filtered; then dividing the urban area range into 128 x 128 grid areas, wherein the size of each area is 200 x 200 m; the setting of the grid area is more beneficial to the analysis efficiency of the data; the grid area of 200 meters can still be matched with the road section information;

then obtaining historical operation data in an operation area, wherein the historical operation data comprises average time from grid i to grid j at the time t, average income of passengers from grid i to grid j at the time t, time for patrolling grid i at the time t, the number of taxis of the passengers (shown as a hot spot diagram of taxi-on positions in different time periods, a diagram of taxi-on position hot spots in a time period of 6-7 points in fig. 2(a), and a diagram of taxi-on position hot spots in a time period of 9-10 points in fig. 2 (b)) found in grid i at the time t, the number of taxis passing through grid i at the time t, the number of passengers taking area j as a destination at the time t, and the like, wherein t is a positive real number, and i and j are natural numbers; basic data for subsequent analysis.

The operating time of all taxis is first calculated to estimate the taxi driver's rental time: operating time of taxi (using T)_busTo represent) comprises two parts, the total passenger carrying time (denoted by T)_driveExpressed in terms of T) and total seek time (in terms of T)_seekRepresented by (a) as shown in the following formula:

T_bus＝T_drive+Ts_eek

the total passenger carrying time is the sum of the passenger carrying time of all transactions of one taxi. The total seek time is equal to the sum of the time between all consecutive transactions. But the calculation of the seek time is more critical because the time difference between two consecutive transactions varies and some drivers may choose to take a break between transactions. Statistical analysis is performed according to historical data, and it can be known that 90% of the guest-searching time is less than 25 minutes, so that 25 minutes is taken as a time threshold herein; data above this threshold will not count into the total seek time, as the taxi driver may choose to rest, refuel, change shifts, etc. during this time.

The income efficiency is as follows: the longer the taxi is in operation, the higher the total revenue will be. But the operating time per day is always fixed (not exceeding 24 hours), so it is more important how to maximize the revenue per unit time within the given time limit. Accordingly, a revenue efficiency concept is defined (by E)_revRepresents): the revenue per unit operating time for the taxi driver is shown as follows:

where M represents the total revenue of the taxi driver, the revenue efficiency formula may be applied to an hour, a day, or even a year.

Definition of successful and unsuccessful taxi drivers: ranking taxi drivers in white class according to income efficiency, and then finding taxi drivers ranked in front of and behind; analysis based on historical data shows that about 80% of taxi drivers have revenue efficiencies between 0.72 and 0.80. Thus taxi drivers ranked 10% of revenue efficiency are defined as successful drivers, and taxi drivers ranked 10% of revenue efficiency are defined as unsuccessful drivers; meanwhile, time periods are also divided to compare the revenue efficiencies of the top 10% taxi drivers and the 10% taxi drivers after the taxi drivers are ranked (as shown in fig. 3 (a)), the driving efficiency (as shown in fig. 3 (b)) and the average search time (as shown in fig. 3 (c)) for comparison; the lighter column in fig. 3 represents the top 10% taxi drivers and the darker column represents the bottom 10% taxi drivers.

The total operation time of the taxi comprises passenger carrying time and passenger searching time, the taxi only has income under the condition of carrying passengers, and the taxi does not have income when searching passengers. The revenue efficiency of a taxi driver is therefore dependent on 1, how much money can be earned in the event of a passenger (passenger efficiency); 2. whether the taxi can find the next passenger faster when the taxi is in idle running (passenger searching efficiency). When inexperienced taxi drivers choose a slower route or enter a congested section of traffic, their passenger and revenue efficiencies are often reduced. First, the following passenger carrying efficiency E is defined_drive:

The best way to improve the income efficiency is to improve the passenger carrying efficiency E_driveWhile reducing the total seek time. Fig. 3(b) compares the passenger carrying efficiency of the top 10% taxi drivers and the bottom 10% taxi drivers, which are about 10% to 13% different. Meanwhile, the average search time of the taxi drivers ranked 10% before and 10% after each time slot is compared, as shown in fig. 3(c), the result shows that the search time can be saved by 25% to 35% by the taxi drivers ranked 10% before compared with the taxi drivers ranked 10% after.

For a taxi, the current state of the taxi needs to be judged, so that a basis can be provided for the next optimization process:

and R3, if the taxi is in the passenger-free state and the duration time of the passenger-free state is less than a set threshold value, determining that the taxi is in a passenger searching state, and optimizing the next operation state of the taxi by adopting the Markov decision process method provided by the invention.

Obviously, the optimal passenger search strategy for increasing taxi revenue per time slot depends on the current state of the taxi and the size of the remaining time window. In the markov decision process model, the taxi's status depends on 3 parameters, the current location (i.e., number of regions, from 1, 2, … to 128 x 128), the current time (the first few minutes in a certain time period), which adjacent region the taxi arrives at the current region, here represented by arrival direction D,

for each state, the direction of arrival is recorded to prevent taxis from being trapped in a situation of cyclic visitor finding in several areas or unlimited on-site waiting. Each region contains 8 neighboring regions around it, so each region corresponds to one direction of arrival, and thus there are 8 directions of arrival to the current region. At the same time define an empty direction

To indicate that the taxi has just been under the passenger in the current zone and therefore does not have any directions to reach. These directions are represented by 10 numbers (0-9), and as shown in fig. 4(a) and 4(b), for any one grid 5, theoretically 8

grids

1, 2, 3, 4, 6, 7, 8, and 9 can reach 5; while for any one mesh 5, there are theoretically 8

meshes

1, 2, 3, 4, 6, 7, 8 and 9 that can be the destination mesh for the next time instant.

In the model, there are 9 actions that can be taken for each empty car in a state. Each action means that the taxi moves from the current area to one of the adjacent 8 areas or stays on the spot to wait.

This process is represented as: a is the same as the A,

the 9 numbers 1-9 are used to indicate directions, as shown in fig. 4 (b). The arrow direction corresponds to the parameter d in the state. The representation of the numbers in d and a is not identical, although the orientation is the same. For example: if the taxi reaches the current area from the left grid area, the direction of arrival d is 4, if the taxi continues to seek passengers to the right, the direction of action a is 6, and the equation d is a-10 for the same direction. The direction not reached (the taxi gets off in the current area and then starts to pick up) is also indicated by 0.

To ensure the accuracy of the MDP model, the following constraints need to be made: 1. taxis cannot enter an invalid grid area (e.g., mountains, rivers, etc.); 2. the adjacent area of the taxi destination must be connected with the current area by a road network. For each grid area number i, all the guest seeking records from the grid area i to 8 adjacent areas are counted. And calculating the probability of seeking passengers from i to j for each area j adjacent to i, and if the probability is lower than a certain threshold value, supposing that the taxi in the historical data rarely seeks passengers in the area j, so that the algorithm does not provide recommendation for seeking passengers in the area j in the grid area i. The probability of the area going to the No. 1 area (upper left) and the No. 3 area (upper right) is 0, so the probability of finding passengers between the area and the adjacent No. 1 area and No. 3 area is very low, and the probability of the area going to the No. 5 area (the area) is also 0 because the taxi is not recommended to stay in place by the algorithm.

Finally, it is also avoided that the taxi driver searches for passengers circularly among a plurality of grid areas or waits in a certain area without limitation, as shown in fig. 4(c), so that the passenger searching efficiency of the taxi driver is reduced. Thus, the method of the present invention requires that the taxi driver take only a subset of the actions called A_allowed(s), but if the taxi driver just alight from the current zone, then the arrival direction is 0, at which time the action that the taxi driver can take may be any direction. When a taxi seeks a passenger in a certain area and does not find a passenger, he must leave the current area. Meanwhile, the direction of the taxi leaving an area cannot form a sharp turn with the direction of the taxi arriving. In particular, if a taxi arrives from direction d at the current grid area l at time t, then he is allowed to take action A_{allowed(l，t，d)}Should make an angle between-90 ° and 90 ° with the direction of arrival, as shown in fig. 4 (d). For example: a taxi coming to the current area from the left d direction (direction 4) then he can take action a (2, 3, 6, 8, 9).

Assume that the current taxi state is s ═ i, t, d. Taxi action a will move the taxi from grid area i to j, then the taxi will roam area j to find the next passenger, and assume that the time to roam the entire area j is t_seek(j) And (3) minutes. Then the following two results will be produced:

(1) taxi patrols t_seek(j) The next passenger is successfully found in region j, and the taxi will reach the destination designated by one passenger (assumed to be k), with the probability P from j to k_dest(j, k). Once the taxi arrives at destination k, the transaction is completed, using t_drive(j, k) as the time it takes for this transaction, and the revenue for this transaction is assumed to be r (j, k), which is the expected revenue for the taxi from j to k. Then the taxi will continue to search for the next passenger again from k. The state of taxi is changed to (k, t + t)_seek(j)+t_drive(j，k)，0)。

(2) If the taxi tours t in the area j_seek(j) After minutes no passenger is found, the taxi must leave the current area to go to the next area to find the passenger.

Assuming that the behavior a taken by the taxi is 6(→), then the state after the taxi tour is finished is s ═ j, t + t_seek(j) 4) (because of taxiFrom the left).

In summary, an empty vehicle is in any state s₀＝(j，t，d)(s₀E S) has a certain possibility to take action a e A_allowed(s₀) To tour the neighboring region j. And has a probability of 1-P_find(j) Slave state s₀Is converted into a state s₁＝(j，t+t_seek(j) 10-a), this state transition is without any benefit. At the same time there will also be a probability P_find(j)×P_dest(j, k), (k ═ 1, 2, … | L |) taxi from state s₀Is converted into a state s₂＝(k，t+t_seek(j)+t_drive(j, k),0), and this state transition would have the benefit of an r (j, k) element.

The goal of the model is to maximize the revenue for the taxi driver for the current time period. The model has a series of end states, that is, the time is up to the end of a certain period, that is, at the 60 th minute. Once this time is reached, the taxi will take no further action. And starting at the first minute of the next time period, and continuing to search for the passenger according to the recommendation.

For each action a, the current time period is denoted by V x (s, a), and in state s, the expected benefit that action a can bring is. V(s) represents the maximum benefit of remaining time available at state s. This process is expressed using the following formula:

where s is a state and a is taxi from l to l_aThe action taken; and recording all path processes established by the formula, and optimizing the income of the taxi.

In the formula, max is a function of solving the maximum value, and V (s, a) is expected yield which can be brought by action a of the taxi in a state s in the current time period; p_find(l) Probability of a taxi being able to find a passenger in grid l, V (l)_a,t+t_seek(l_a) 10-a) T + t has been spent for the taxi_seek(l_a) Time and in the current grid area l_aIf no passenger is found, taking (10-a) as the recommended expected value of the driving direction; v (k, t + t)_seek(j)+t_drive(j, k),0) represents the expected value of finding a passenger in the j region, where t + t_seek(j)+t_drive(j, k) represents finding a passenger in the j area to go to the k area, when the time spent searching and the time spent from the j road section to the k road section to be spent are about to be spent; t is the current time, t_seek(l_a) For touring target grid area l_aRequired time, P_dest(j, k) is the probability from grid area j to grid area k, r (j, k) is the expected revenue for taxis from grid area j to grid area k, t_drive(j, k) is the time it takes for the taxi to go from grid area j to grid area k; a is an action taken by the vacated taxi, a belongs to A, and A is an action set which can be taken by the vacated taxi; a is a natural number of 1-9.

Probability P that taxi can find passenger in grid i_find(i) The method comprises the following steps The number of taxis for which passengers are found in zone i is compared with the total number of taxis for which the passenger passes through zone i (including the number n of taxis for which passengers are found)_findAnd the number n of taxis for which no passenger is found_passSum) is the probability that the taxi can find the passenger in region i, for example: a road segment passes 100 vehicles, of which 50 vehicles find passengers, and the probability of finding a passenger on that road segment is 50%. Since only data for getting on or off the taxi in an area is adopted, a path to be taken from the taxi to the next passenger needs to be estimated. For each passenger seeking, all the paths passing from the starting point (taxi getting-off area) to the terminal point (next passenger carrying area) are calculated by using the existing algorithm. And then mapping all the paths into the grid area, thereby obtaining the number of taxis passing by each area in each time period. The probability expression for finding a passenger is shown as follows:

for theProbability P from mesh region i to mesh region j_dest(i, j): to estimate all region destination probability parameters for each time segment, a 16384 × 16484 matrix W of the number of each grid region to all other regions is calculated from the historical data, with the parameter W in the matrix_ijRepresenting the number of taxis from grid area i to area j for the corresponding time period. W is to be_ijComparing the total number of cars on zone i, i.e. the probability P from zone i to zone j_dest。

Expected revenue r (j, k) for taxis from grid area j to grid area k: the average income between any two areas (from the getting-on area to the getting-off area) in each time period is counted as the expected income between the areas.

Time t spent by taxi from grid area j to grid area k_drive(j, k): assuming that the gains obtained from one region to other regions are the same for the first 10% taxi drivers and the second 10% taxi drivers, then considering the average driving efficiency of the first 10% taxi drivers and the second 10% taxi drivers, respectively calculating the ratio of the total income and the total driving time of the two groups of taxi drivers as the driving efficiency. By using driving strategies, the time t spent in each transaction can be estimated by the transaction revenue_drive(i,j)＝r(i,j)/E_drive。

And finally, solving the current function by adopting a dynamic planning method so as to optimize the next operation route of the taxi.

Finally, the optimization effect of the method is verified through a simulation calculation process:

in the historical taxi position data, after 10% of taxi drivers ranked at the front and 10% of taxi drivers ranked at the back are taken off, based on the method, the optimal taxi-seeking area is recommended to the taxi driver, and the taxi-seeking process is simulated, namely after the taxi driver takes off, according to the time of taking off and the getting-off place, the taxi driver is recommended to the next area by the algorithm to seek passengers, and when the taxi driver reaches the recommended area, whether the taxi driver can find the passenger in the current area is judged by utilizing the seeking probability of each area obtained in the previous step. If the passenger can not be found, continuously recommending the taxi driver to go to the next area for finding the passenger, if the passenger is found, selecting a destination area to give the taxi driver according to the probability distribution of the passenger destinations getting on the area in the historical data, simultaneously obtaining corresponding benefits by the taxi driver, and continuously recommending the taxi driver after getting off the taxi again until the time reaches the end of a time period.

According to the above, it is assumed that taxi drivers have different driving efficiencies, so that the driving efficiencies of taxies 10% before and 10% after the ranking are used in the simulation process to determine the time T taken by the taxi drivers to get to or get from two areas_drive. The results of fig. 5 below were obtained by simulating 6000 times the top 10% and the bottom 10% taxi drivers, respectively. Fig. 5(a) -5 (b) show revenue efficiency profiles for top 10% ranked taxi drivers and revenue efficiency profiles based on simulation results of top 10% ranked driving efficiency. For the top 10% of taxi drivers, their average revenue efficiency increased from 0.828 yuan per minute to 0.89 yuan per minute, with an increase of 7.6%. Fig. 5(c) -5 (d) show revenue efficiency profiles of 10% taxi drivers after ranking and revenue efficiency profiles based on simulation results of 10% driving efficiency after ranking. Their average revenue efficiency increased from 0.758 yuan per minute to 0.82 yuan per minute with a 9.8% increase.

Simulations were performed for all day periods using the same method and the results were compared. Fig. 6(a) compares the top 10% taxi drivers with the revenue efficiency comparison based on the driving efficiency of the top 10% taxi drivers. The method (indicated as MDP in the figure) provided by the invention is improved by 3.6% to 12% compared with the taxi driver with the top 10% ranking. Fig. 6(b) compares the revenue efficiency comparison of the 10% taxi drivers after ranking and the driving efficiency based on the 10% taxi drivers after ranking. The algorithm provided by the invention is improved by 4.2% to 15.5% compared with the taxi drivers with the top 10% of the ranking. As can be seen from the figure, the recommendation effect of the algorithm in some peak periods in the morning and afternoon is improved more remarkably, because more reference data are available at the moment, the researched parameters are more accurate, and the recommendation is more accurate. Fig. 6(c) compares the revenue efficiency comparison of the driving efficiency of the top 10% taxi drivers and the rear 10% taxi drivers. By comparison, after the method is used for recommending, the income efficiency of 10% of taxi drivers after ranking can reach the income level of 10% of taxi drivers before ranking, and the effectiveness of the method is further verified.

Finally, the recommendations of the method of the invention were compared to a similar Maximum Net Profit (MNP) algorithm. In the MNP method, given a time period and a getting-off position, 5 road sections with the largest expected profit are recommended to be provided for a car rental driver to search for passengers, and if the car rental driver does not find the passengers, the MNP continuously recommends 5 results at the last ending road section to be provided for the car rental driver.

The grids are adopted to replace road sections for comparison, under the same parameter setting, the two methods are simulated 6000 times in each time period, and the simulation result in fig. 6(d) shows that the recommendation method based on the MDP is obviously improved in each time period compared with the recommendation method based on the MNP, and the maximum improvement rate is 8.4%.

The patent of the invention obtains the funding of the key research and development plan of the science and technology hall in Hunan province (project number: 2017GK 2272).

Claims

1. The taxi running route optimization method for multiple passenger mileage comprises the following steps:

s1, dividing an operation area of a taxi into a plurality of grids, and simultaneously acquiring historical operation data in the operation area; the size of the divided grid area can be matched with road section information in the operation area of the taxi; the historical operation data comprises average time from grid i to grid j at time t, average income for transporting passengers from grid i to grid j at time t, time for traveling to grid i at time t, the number of taxis of passengers found in grid i at time t, the number of taxis passing through grid i at time t, and the number of passengers taking region j as a destination at time t, wherein t is a positive real number, and i and j are natural numbers;

s2, positioning the current position and operation state of the taxi;

s3, optimizing the next operation route of the taxi according to the current position and the operation state of the taxi acquired in the step S2, so that the income efficiency optimization of the taxi is completed; the step of optimizing the next operation route of the taxi comprises the following steps:

A. obtaining the current operation state of a taxi;

B. b, optimizing the next operation route of the taxi by adopting a Markov decision process method according to the operation state of the taxi acquired in the step A;

the optimization of the next operation route of the taxi specifically adopts the following rules:

r1, if the taxi is in a state of transporting passengers, optimizing a next-step operation route of the taxi by adopting a path optimization algorithm;

in the formula, max is a function of solving the maximum value, and V (s, a) is expected yield which can be brought by action a of the taxi in a state s in the current time period; p_find(l) Probability of a taxi being able to find a passenger in grid l, V (l)_a,t+t_seek(l_a) 10-a) T + t has been spent for the taxi_seek(l_a) Time and in the current grid area l_aIn the case where no passenger is found, the recommended period with (10-a) as the driving directionA desired value; v (k, t + t)_seek(j)+t_drive(j, k),0) represents the expected value of finding a passenger in the j region, where t + t_seek(j)+t_drive(j, k) represents finding a passenger in the j area to go to the k area, when the time spent searching and the time spent from the j road section to the k road section to be spent are about to be spent; t is the current time, t_seek(l_a) For touring target grid area l_aRequired time, P_dest(j, k) is the probability from grid area j to grid area k, r (j, k) is the expected revenue for taxis from grid area j to grid area k, t_drive(j, k) is the time it takes for the taxi to go from grid area j to grid area k; a is an action taken by the vacated taxi, a belongs to A, and A is an action set which can be taken by the vacated taxi; a is a natural number of 1-9, wherein 1 represents that the next action of the taxi is driving from the current position to the southwest direction, 2 represents that the next action of the taxi is driving from the current position to the southwest direction, 3 represents that the next action of the taxi is driving from the current position to the southeast direction, 4 represents that the next action of the taxi is driving from the current position to the southwest direction, 5 represents that the next action of the taxi is staying at the current position, 6 represents that the next action of the taxi is driving from the current position to the southwest direction, 7 represents that the next action of the taxi is driving from the current position to the northwest direction, 8 represents that the next action of the taxi is driving from the current position to the northwest direction, and 9 represents that the next action of the taxi is driving from the current position to the northeast direction.

2. The method for optimizing taxi running route based on multiple passenger mileage as claimed in claim 1, wherein the next step of taxi running route optimization according to rule R3 is specifically to solve the current function by using a dynamic planning method, so as to optimize the next step of taxi running route.