PreprintPDF Available

Abstract and Figures

Matching trip requests and available drivers efficiently is considered a central operational problem for ride-hailing platforms. A widely adopted matching strategy is to accumulate a batch of potential passenger-driver matches and solve bipartite matching problems repeatedly. The efficiency of matching can be improved substantially if the matching is delayed by adaptively adjusting the matching time interval. The optimal delayed matching is subject to the trade-off between the delay penalty and the reduced wait cost and is dependent on the system's supply and demand states. Searching for the optimal delayed matching policy is challenging, as the current policy is compounded with past actions. To this end, we tailor a family of reinforcement learning-based methods to overcome the curse of dimensionality and sparse reward issues. In addition, this work provides a solution to spatial partitioning balance between the state representation error and the optimality gap of asynchronous matching. Lastly, we examine the proposed methods with real-world taxi trajectory data and garner managerial insights into the general delayed matching policies. The focus of this work is single-ride service due to limited access to shared ride data, while the general framework can be extended to the setting with a ride-pooling component.
Content may be subject to copyright.
Optimizing matching time intervals for ride-hailing services using
reinforcement learning
Guoyang Qina, Qi Luob, Yafeng Yinc,, Jian Suna, Jieping Yed
aKey Laboratory of Road and Traffic Engineering of the State Ministry of Education, Tongji University,
Shanghai 201804, China
bDepartment of Industrial Engineering, Clemson University, Clemson, SC 29631, United States
cDepartment of Civil and Environmental Engineering, University of Michigan, Ann Arbor, MI 48109,
United States
dDiDi AI Labs, Didi Chuxing, Beijing, 100085, China
Matching trip requests and available drivers efficiently is considered a central operational
problem for ride-hailing platforms. A widely adopted matching strategy is to accumulate a
batch of potential passenger-driver matches and solve bipartite matching problems repeat-
edly. The efficiency of matching can be improved substantially if the matching is delayed by
adaptively adjusting the matching time interval. The optimal delayed matching is subject
to the trade-off between the delay penalty and the reduced wait cost and is dependent on
the system’s supply and demand states. Searching for the optimal delayed matching policy
is challenging, as the current policy is compounded with past actions. To this end, we tailor
a family of reinforcement learning-based methods to overcome the curse of dimensionality
and sparse reward issues. In addition, this work provides a solution to spatial partition-
ing balance between the state representation error and the optimality gap of asynchronous
matching. Lastly, we examine the proposed methods with real-world taxi trajectory data
and garner managerial insights into the general delayed matching policies. The focus of
this work is single-ride service due to limited access to shared ride data, while the general
framework can be extended to the setting with a ride-pooling component.
Keywords: Ride-hailing service, online matching, reinforcement learning, policy gradient
1. Introduction
Assigning trip requests to available vehicles is one of the most fundamental operational
problems for ride-hailing platforms, who act as a mediator that sequentially matches supply
(available drivers) and demand (pending trip requests) (Alonso-Mora et al.,2017;Wang
and Yang,2019). The common objective is to maximize the system revenue or minimize
passengers’ average wait time, both of which necessitate devising efficient matching strategies
(Zha et al.,2016).
Corresponding author, email:
Preprint submitted to Transp. Res. Part C Emerg. Technol. (accepted version) June 19, 2021
Qin et al. You can read the published version here Transp. Res. Part C
A matching strategy widely adopted by ride-hailing platforms is to formulate and solve a
bipartite matching problem between available vehicles and trip requests (Xu et al.,2018). As
illustrated in Fig. 1, intentionally controlling the matching time interval to delay the match-
ing of passengers and drivers can adjust the buffer size of potential matches and improve
the operational efficiency, because drivers may find a better (i.e., closer, more profitable, or
more compatible) match when demand accumulates and vice versa. As a result, an obvious
trade-off arises when extending this matching time interval. On the one hand, the outcome
of matching is improved with the increasing length of passengers’ and drivers’ queues; on
the other hand, waiting induces penalty costs for each driver and passenger in queues and, in
worst-case scenarios, leads to abandonment. This type of rewarding delay effect is prevalent
in online matching markets. Various online platforms, such as dating applications or organ
donation systems, have adopted this kind of delayed matching strategy (Azar et al.,2017).
Figure 1: An appropriate matching delay can increase the passenger-driver buffer size and improve the opera-
tional efficiency. (a) Passengers and drivers may suffer rather long pickup wait times under an instantaneous
matching scheme, (b) moderately delaying the matching can significantly reduce the pickup wait times.
Upon closer inspection into the dynamic matching process, one may see that the appro-
priate matching time interval depends on the ride-hailing system’s state, i.e., spatiotemporal
distributions of supply and demand. One of this work’s main findings is that, when ei-
ther the supply or demand is scarce, the platform should use instantaneous matching to
maximize the system throughput. However, when the supply and demand are relatively
balanced, strategically controlling the matching time interval is superior. There has been
a significant research interest in the optimal control policy for the latter case (Azar et al.,
2017;Ashlagi et al.,2018;Yang et al.,2020). Limitations in prior studies include requiring
parametric models to obtain analytical solutions or mandating unrealistic assumptions on
the system’s stationarity. Nevertheless, these studies’ numerical experiments show that the
optimal matching time interval is distinctly sensitive to these assumptions. Hence, these
model-based approaches are not implementable in complex ride-hailing applications.
Due to the state space’s large dimensionality, including vehicle locations, demand fore-
cast, and passenger characteristics, obtaining a pragmatic and effective nonparametric con-
trol policy for the matching time interval remains an open question. Moreover, when the
Qin et al. You can read the published version here Transp. Res. Part C
system uses a fixed matching time interval, there is no prior information on the ride-hailing
system dynamics under alternative time intervals. Reinforcement learning (RL) is therefore
an obvious choice to handle this task. This work aims to develop an efficient, data-driven
approach to find the optimal matching time intervals. The focus of this work is single-
ride service due to limited access to shared ride data, while the general framework can be
extended to the setting with a ride-pooling component.
The main contributions of this work include:
a. Proposing a dedicated reward signal for passenger-driver matching with delay, which
addresses the sparse reward issue related to a long matching time interval.
b. Providing a remedy to the spatial partitioning trade-off between the state represen-
tation error and the optimality gap of local matching. This trade-off arises when we
partition the area and apply asynchronous policies over those grids.
c. Presenting managerial insights for an optimized matching delay based on experiments
with real-world data. The RL algorithm identifies the optimized delayed matching
policy delay under medium driver occupancy.
The remainder of this paper is organized as follows. We first review the related literature
in Section 2, then provide a concrete formulation of the sequential decision problem in Section
3. We investigate three RL algorithms in Section 4as each one is a building block for the
more advanced one. Finally, we show the results in a real-world case study in Section 5and
draw the final conclusion in Section 6.
2. Literature review
Online matching strategies are characterized by two spatiotemporal variables in general:
the matching time interval and the matching radius (Yang et al.,2020). The platforms
dynamically adjust these variables to improve the system performance measures such as the
gross revenue or cumulative system throughput. The effectiveness of matching strategies
is critical for the platform’s profitability and the quality of service perception of customers
(Wang and Yang,2019). Moreover, due to the negative externalities induced by empty-car
cruising (Luo et al.,2019), developing advanced matching strategies ensures the sustainability
of ride-hailing services.
A number of studies in the literature obtained analytical results for optimal matching
strategies assuming that the system dynamics is known. Most matching strategies devel-
oped in the past consider instantaneous matching, i.e., the platform assigns trip requests to
drivers upon arrival. For example, considering instantaneous matching, Xu et al. (2020b)
investigated the impact of matching radius on the efficiency of a ride-hailing system and the-
oretically proved that adaptively adjusting matching radius can avoid throughput loss due
to ”wild goose chases.” Aouad and Sarita¸c (2020) modeled the dynamic matching problem
as a Markov decision process with Poisson arrivals and developed a 3-approximation algo-
rithm for the cost-minimization with uniform abandonment rates. They also showed that the
batched dispatching policy’s performance can be arbitrarily bad in the cost-minimization set-
ting, even with the optimally tuned batching intervals. ¨
Ozkan and Ward (2020) studied the
optimal dynamic matching policy in a ridesharing market modeled as a queueing network.
Since solving the exact dynamic program for the joint pricing and matching problem with
Qin et al. You can read the published version here Transp. Res. Part C
endogenous supply and demand is intractable, they proposed a continuous linear-program-
based forward-looking policies to obtain the upper bounds of the original optimization. The
value of delayed matching is obvious for two-sided online service systems (Azar et al.,2017).
As the thickness of the market increases over time, passengers and drivers are provided with
superior candidates for potential matching with a cost of delay. Ashlagi et al. (2018) studied
the near-optimal matching strategies with either random or adversarial arrivals of passen-
gers and drivers. The approximation ratio for the random case is constant, which means
the algorithm’s efficiency is unchanged for large networks. However, challenges remain as
the deadlines for matching were given in their work. Yang et al. (2020) separated the opti-
mization of the matching radius and matching time interval intending to maximize the net
benefit per unit time. They obtained the optimal matching time interval only under the
excess supply scenario and ignoring passengers’ abandonment behaviors. Yan et al. (2020)
studied the jointed matching and pricing problem in a batched matching setting. The wait
time and matching time trade-off was identified in the equilibrium analysis. In addition,
high carry-over demand or supply impeded their analysis. Most of these analytical results
focused on the non-ridepooling setting.
Practitioners are more concerned about developing purely data-driven approaches for
passenger-driver matching in ride-hailing platforms. Since online matching is a sequential
decision under uncertainty, reinforcement learning (RL) is a powerful technique to adaptively
improve the matching policy (Li et al.,2019;Al-Abbasi et al.,2019;Lin et al.,2018;Ke
et al.,2020;Mao et al.,2020). The main technical challenge is the curse of dimensionality
due to the large state space – the system contains a large number of cruising drivers and
waiting passengers. The benefit of matching a specific passenger-driver pair, i.e., the reward
function, is determined by the geographical information and characteristics. Li et al. (2019)
approximated the average behavior of drivers by mean field theory. Al-Abbasi et al. (2019)
used a distributed algorithm to accelerate the search of near-optimal vehicle dispatching
policies. Mao et al. (2020) studied the vehicle dispatching policy using an Actor-Critic
method. These studies require instantaneous matching to avoid many of the difficult reward
design problems. Ke et al. (2020) adopted an RL-based method for the delayed passenger-
driver matching problem. They employed a two-stage framework that incorporated a multi-
agent RL method at the first stage and a bipartite matching at the second stage. Their work
addressed several technical issues that were unresolved in previous studies, such as learning
with delayed reward and value error reduction. By combining a queueing-type model with
Actor-Critic with experience replay (ACER), we improve the practicability and stability of
RL approaches for passenger-driver matching with delay.
3. Problem description and preliminaries
This section first formulates the delayed matching problem as a Markov decision process
(MDP). We then show the existence of optimal matching time interval due to the delayed
matching trade-off. Lastly, we describe how to create a simulator for agents to learn optimal
policies with streaming information from a ride-hailing system, which is of particular interest
to the transportation field.
Qin et al. You can read the published version here Transp. Res. Part C
3.1. Delayed matching problem in ride-hailing
In the process of online passenger-driver matchmaking, a matchmaker makes recurring
two-stage decisions through a sequence of discrete-time steps t∈ {0,1,· · · , T }.Tcan be
either finite or infinite. Given a partitioned area and time step t, the first-stage decision is
about whether to hold or match the batched passengers and drivers in each grid; this decision
can be made asynchronously in different grids. The second-stage decision is to optimize the
matching of the batched passengers and drivers in the combined grids where the matchmaker
decides to match. The optimization objective of these sequential decisions is to maximize
the total wait reward 1The notation used throughout this work is summarized in Table A.2
in Appendix A.
This sequential decision-making problem can be elaborated as a Markov decision process
(MDP). At each time step t, the matchmaker first observes the grid-based system state
matrix S
S(t) by gathering information, including currently batched passenger trip details and
idle driver locations, as well as predictions about the future supply and demand in each
grid. The matchmaker then follows a policy function π(A
A(t) = a|S
S(t) = s) to make the first-
stage decision. All passengers and drivers within the grids with action A
A(t) = 0 (“hold”)
will be held in the buffer for one more time step. Otherwise, all passengers and drivers
within the grids with action A
A(t) = 1 (“match”) will be globally matched in the second-
stage decision. Assuming that each driver can take at most one passenger request per match
(i.e., not considering ride-pooling), we make the second-stage decision by solving a min-cost
bipartite matching problem (Williamson and Shmoys,2011). At the end of these two-stage
decisions, unless abandoning the queue, unmatched drivers or passengers will circulate back
to the buffer and new arrivals will be admitted into the buffer. Further, the system state
will transition to S
S(t+ 1) accordingly.
If we define the wait reward incurred as R
R(t) = r(S
S(t+ 1)) and a state tran-
sition probability p(S
S(t+ 1)|S
A(t)), the expected total wait reward starting from tcan
be expressed as Vπ(S
S(t)) = Eπ[P
k=t+1 γkt1R
R(k)], where γis a discount factor. Then the
matchmaker’s objective is to find an optimal policy πto maximize the expected total reward
S(0)) = Eπ[
γt1R(t)] (1)
S(0)) X
S(0), a, s0) + Vπ(s0)]p(s0|S
S(0), a),
where the second summation term can be denoted as an action value function qπ(s, a),
qπ(s, a) := X
[r(s, a, s0) + Vπ(s0)]p(s0|s, a).(2)
1To align with the reinforcement learning terminology, we use the term “wait reward”, the negated wait
cost, and its derivatives such as reward signal, reward design, and sparse reward issue when dealing with the
RL modelling, while in other cases, we use the more intuitive term “wait cost”. To this end, maximizing the
wait reward and minimizing the wait cost become the equivalent objectives of the optimization.
Qin et al. You can read the published version here Transp. Res. Part C
If all the aforementioned functions and their parameters are known,πcan be exactly
solved, regardless of the large state space of S
S(t). However, the formulation above deems
intractable in practice. The reasons are two-fold:
a. As we have no prior knowledge of either the transition probability p(s0|s, a) or the
reward function r(s, a, s0), exact solutions are ruled out for this problem.
b. Aggregation of the original state space matrix S
S(t) to some smaller space produces
partially observed state, which may violate the Markov assumption in MDP2.
These two challenges suggest that finding an optimal on/off decision policy for an online
passenger-driver matchmaking system is by no means trivial. Next sections will elaborate
on how they are approached.
3.2. Preliminaries on the matching delay trade-offs
This subsection explains how the matching time interval affects the matching wait time
and the pickup wait time in a ride-hailing system. We introduce a deque model to approx-
imate the matching process (Xu et al.,2020b). Due to deque’s conciseness in theory and
descriptiveness in practice, we can derive its performance metrics accordingly.
Without loss of generality, we assume passengers and idle drivers arrive by two inde-
pendent stationary Poisson processes (arrival rate λPand λDresp.). Their locations are
uniformly distributed within a square area (size of d×d). Note that the choice of demand
and supply models does not affect the proposed RL algorithm in the following section. Newly
arrived passengers and drivers will balk or renege the queue if the expected wait time exceeds
a threshold tmax. Upon their arrival, both passengers and drivers will first be batched in a
buffer pending matching. Pickup wait times are estimated by the Manhattan distance be-
tween passengers and drivers divided by the average pickup speed v. Unmatched passengers
or drivers will circulate back to the buffer and be carried over to the next matching time win-
dow. A fixed matching time interval strategy matches passengers and drivers in a bipartite
graph in the second stage at each of a sequence of discrete-time steps t∈ {t, 2∆t, 3∆t, · · · }.
This method is suitable if deque arrivals are stationary, in which the delayed matching prob-
lem is equivalent to finding an optimal fixed ∆t. If the arrivals are non-stationary, optimizing
adaptive t’s becomes substantially more challenging.
Fig. 2illustrates the trade-off curves derived via simulation from the deque model under
scenarios with different supply (λD) and demand (λP). As the matching time interval ∆t
extends, the matching wait time increases in a approximately linear fashion, as shown in Fig.
2(a), while the pickup wait time curves in Fig. 2(b) descend more steeply. In particular, the
scenarios with balanced supply and demand have the steepest U-shape trade-off curves in Fig.
2(c). These trade-off curves empirically manifest the considerable potential of implementing
delayed matching in a ride-hailing platform.
Contour plots in Fig. 3further depict the minimum expected total wait time and its
corresponding optimal ∆t, which can be regarded as a value function and a policy function
under stationary arrivals. Fig. 3(a) shows that the minimum expected total wait time
2The transformation of state space may retain the Markovian properties under conditions, and interested
readers are referred to Proposition 3 in Kim and Smith (1995).
Qin et al. You can read the published version here Transp. Res. Part C
Matching wait time
I J4.0 -
4.5 -
5.0 -
5.5 -
6.0 -
Pickup wait time Total wait time
0 50 100 150
Fixed matching time interval (s)
0 50 100 150
Fixed matching time interval (s)
0 50 100 150
Fixed matching time interval (s)
Figure 2: Change of passenger wait time as the matching time interval is extended in a deque model.
Minimum expected total wait time (s)
Optimal fixed matching time interval (s)
5.5 5.5
~ 5.0 ~ 5.0
4.5 4.5
4.0 ~ ~ L_...l........-...l..,_
4.0 4.5 5.0 5.5 6.0 4.0
4.5 5.0
5.5 6.0
Figure 3: Minimum expected total wait time and its corresponding optimal fixed matching time interval.
increases as the gap between λPand λDbroadens, while the minimum wait times are similar
if the difference between λPand λDis held constant. Fig. 3(b) shows that the optimal
matching intervals must be substantially prolonged if λPand λDare balanced, otherwise
they should have a shorter delay. The contour patterns clearly corroborate the necessity of
delayed matching, especially when λPλD.
In the next subsection, we will generalize the deque model to a ride-hailing simulator as
a dock for accommodating non-stationary and spatially-dependent arrivals. The rest of the
current work is underpinned by this simulator.
3.3. A queueing-based ride-hailing simulator
Qin et al. You can read the published version here Transp. Res. Part C
This section adopts a reinforcement learning approach that approximates the transition
dynamics in the MDP to resolve the first challenge in the original formulation. A ride-
hailing system simulator enabling the RL approach to sample sequences of states, actions,
and matching rewards is needed. To be specific, a ride-hailing simulator provides a control-
lable interface that connects RL agents to an approximate real-world ride-hailing system.
On the real-world side, the simulator is required to simulate trip request collection, delayed
matching, and delivery of passengers. On the learning agents’ side, the simulator is required
to generate observations O
O(t) that characterize the ride-hailing system’s state S
S(t) and pro-
vide informative reward signals. In this regard, we decompose the simulator creation task
into two parts: (a) developing an online ride-hailing service simulator and (b) designing
observations and reward signals that assist in agent learning.
3.3.1. Creating an interactive simulator for RL
Our first attempt is to reduce the environment complexity by modeling the arrival of the
passenger requests and idle drivers as a multi-server queueing process, as shown in Fig. 4.
vehicle servers
bipartite matchingaction taking
retain holding
Figure 4: Workflows of the online ride-hailing service simulator.
Arriving passenger requests are first admitted into a buffer. Apropos of action matrix
A(t), actions will be taken separately in two subareas (illustrated in Fig. 5), where the
subarea containing all the grids whose A(t) = 0 is denoted as S0, and the subarea containing
all the grids whose A(t) = 1 as S1. The buffer will retain holding requests in the subarea
S0while releasing the batched requests in the subarea S1to the bipartite matching module.
Note that partitioning the area and expanding the one-value action to a mixed-action matrix
allows the agents to accommodate spatially dependent arrivals. Once the bipartite matching
module receives requests, it will pull all idle drivers and execute a global bipartite matching
between these two sides in S1. For simplicity, we estimate the edge cost by the Manhattan
distance between passengers and drivers divided by the average pickup speed vand solve for
the minimum pickup wait time. Those requests matched are forwarded to the corresponding
vehicle servers module, while those unmatched requests return to the buffer.
To be specific, the second stage bipartite matching is solved as a linear sum assignment
Qin et al. You can read the published version here Transp. Res. Part C
grid-based actions
bipartite matching
01001 1
10110 0
10110 0
11110 0
10101 1
Figure 5: Grid-based actions and subarea-based matching in the ride-hailing simulator.
problem (LSAP). LSAP is modelled as
Tij xij (3)
xij = 1 (i= 1,2,· · · , n)
xij = 1 (j= 1,2,· · · , n)
xi,j ∈ {0,1}(i, j = 1,2,· · · , n),
where Tij is the pickup time between driver i(1,2,· · · , ND) and passenger j(1,2,· · · , NP)
estimated by their Manhattan distance divided by the average pickup speed v. In case
ND6=NP,Tij will be filled into an n×nmatrix with a very large value M, where
n= max(ND, NP). Namely, Tij =Mfor i(ND+ 1,· · · , n) or j(NP+ 1,· · · , n). In
the ridepooling setting, the second stage problem is to solve a general assignment problem
(Alonso-Mora et al.,2017), and the queueing model needs to be modified to be a match-
ing queue (Cao et al.,2020). Note that these extensions do not affect the general learning
Simulating idle drivers’ repositioning decisions between these grids in a multi-server queue
is critical for the accuracy of the simulator. This work assumes that drivers stay in the same
grid, since each grid’ size is relatively large compared to the idle time. In addition, the
simulator uses a fixed patience time for passengers throughout this study and assumes that
drivers always stay active in the system. These features can be easily extended and embodied
in the simulator with new data access. The key motivation for developing such a ride-hailing
Qin et al. You can read the published version here Transp. Res. Part C
simulator is to provide an interface for agents to evaluate and improve their policy π(a|s),
which can been seen as a special case of the model-based RL (Sutton and Barto,2018).
4. Methodology
This section first describes the design of states and reward signals of the ride-hailing
environment, which provide essential information for RL agents to improve their policies. It
then introduces a family of state-of-the-art RL algorithms called policy gradient methods.
These methods take advantage of the special structure of the delayed matching problem
and navigate the sampled trajectories to approximate the value and policy functions. To
further improve the stability of the algorithm in a dynamic ride-hailing system, we introduce
variants of the Actor-Critic method to solve for the near-optimal delayed matching policy
with streaming real-world data.
4.1. Designing observable states and reward signals of the ride-hailing environment
4.1.1. Observable states
A state can be treated as a signal conveying information about the environment at a
particular time (Sutton and Barto,2018). In a ride-hailing environment, a complete sense
of “how the environment is” should include full information consisting of, but not limited
to, timestamp and origin-destination (OD) of all the batched requests, current locations
of drivers, and predictions about the future supply and demand. However, the complete
state space of the original ride-hailing system is enormous and sampling from it efficiently
becomes impossible. Inspired by the analytical models built in the previous literature (Yang
et al.,2020), we abstract aggregated variables from the ride-hailing simulator as partial
observations of the environment. The observations are concatenated into a tuple as
O(t) = {(NP(i, t), ND(i, t), λP(i, t), λD(i, t)) |iS},(4)
where NP(i, t) (resp., ND(i, t)) is the number of batched passenger requests (resp., idle
drivers) in grid iat time step t, and λP(i, t) (resp. λD(i, t)) is the estimated arrival rate
of passenger requests (resp., idle drivers) in grid ifrom tonward.
Aggregating the complete state space S
S(t) into the observation O
O(t) will generalize the
previously formulated MDP to a partially observable Markov decision process (POMDP),
which echos the second challenge aforementioned. We can assume an unknown distribution
S(t)) linking the complete state S
S(t) and a partial state O
O(t) observed from the deque
model. The matchmaker can re-establish the value function and optimization objective in
this setting by replacing S
S(t) in the MDP setting with O
O(t). In addition, stochastic policies
are enforced to obtain a higher average reward than a deterministic policy (Jaakkola et al.,
Of all the variables in the observation, λP(i, t) is the only variable that is not directly
collectable from the simulator. The learner uses either the average recent arrival rates as a
proxy or a more sophisticated predictor to estimate future demand. In contrast, λD(i, t) is
easy to obtain as the simulator continues to track vehicle locations in real-time. The lost
information of exact locations of vehicles and passengers only affects the first-stage decision
A(t). At the second stage, this information has been considered in the edge cost in the
Qin et al. You can read the published version here Transp. Res. Part C
bipartite matching and included in the reward function. In summary, we reduce the state
space of S
S(t) significantly by transferring partial information noise in O
O(t) of the rewarding
With observable queueing-based states, the state transition dynamics, i.e., how O
should be updated after taking an action, is straightforward:
(NP(i, t) = NP(i, t 1) + nP(i, t 1 : t)
ND(i, t) = ND(i, t 1) + nD(i, t 1 : t),iS0(hold) (5)
NP(i, t) = NP(i, t 1) + nP(i, t 1 : t)
ND(i, t) = ND(i, t 1) + nD(i, t 1 : t)
NM(i, t) = min (NP(i, t), ND(i, t))
NP(i, t) = NP(i, t)NM(i, t)
ND(i, t) = ND(i, t)NM(i, t)
,iS1(match) (6)
where nP(i, t 1 : t) and nD(i, t 1 : t) are the number of trip requests and idle drivers
arriving in grid ibetween t1 and t, respectively. NP(i, t) and ND(i, t) are the number of
requests and drivers updated before being matched at t, as they may change during the time
step. Meanwhile, λP(i, t) and λD(i, t) are continuously re-estimated from historical passenger
arrivals and real-time vehicle locations accordingly.
4.1.2. Reward signals
A reward signal is the environment’s valuation of an implemented action A
A(t). In the
context of supervised learning, this reward signal is required to instruct what action the
agents should take; however it is not required in RL. To this end, a major advantage of RL
distinguishing itself from its counterpart is that the instruction for optimal actions is not
Nevertheless, the reward signals still must be carefully designed, especially when the
environment faces sparse rewards in this setting (i.e., the reward is 0 most of the time in
delayed matching). However, a na¨ıve reward signal will remain unknown until a match is
executed and the total wait cost for all served passengers is determined. In this regard, the
buffer that continues to hold upcoming requests receives zero rewards. This sparse reward
issue makes it extremely difficult for RL to relate a long sequence of actions to a distant
future reward (Hare,2019). In this ride-hailing environment, the agents fed with this na¨ıve
reward are stuck taking a= 0, since the zero reward of taking a= 0 is greater than any
negative rewards incurred by taking a= 1.
To tackle this challenge, we try to decompose this one-shot reward happening at a certain
time into each time step within the passengers’ queueing lifetime. Since the total matching
wait time is in itself a summation of the wait time at each time step, it is straightforward to
delineate its increment as a matching wait reward:
Rm(t) = X
[NP(i, t 1) + τm(nP(i, t 1 : t))] ,(7)
where τm(·) computes the matching wait time of the newly-arrived requests in grid ifrom
t1 to t. In particular, if the arrival of passengers is assumed to be a Poisson process,
Qin et al. You can read the published version here Transp. Res. Part C
τm(nP(i, t 1 : t)) nP(i, t 1 : t)/2.
However, the pickup wait time is not a summation per se; its decomposition is less
straightforward. To resolve this issue, we devise an artifice so that these increments add up
to the final pickup wait time in the matching process. As shown in Fig. 6, the intuition is
to decompose the one-shot pickup wait time τp(t) to a summation of step-wise incremental
pickup wait times, i.e., τp(1),· · · , τp(t1) τp(t2), τp(t)τp(t1).
t-11 ...
Δ pickup wait = τp(1)
Pooled passengers
one-shot pickup wait time
summation of Δ pickup wait
decomposing pickup wait
Figure 6: Decomposition of the one-shot pickup wait time τp(t) to a summation of step-wise incremental
pickup wait times.
In light of this decomposition, we define the step reward mathematically as:
Rp(S, t) = ([τp(NP(S0, t), ND(S0, t)) τp(NP(S0, t 1), ND(S0, t 1))] ,S=S0
τp(NP(S1, t), ND(S1, t)) ,S=S1
where τp(NP(S0,0), ND(S0,0)) := 0, and
NP(S0, t) = X
NP(i, t), NP(S1, t) = X
NP(i, t),
ND(S0, t) = X
ND(i, t), ND(S1, t) = X
ND(i, t).
Note that the difference in Eq. 8is the net increase of the pickup wait reward if holding
(A(S0, t) = 0) requests for one more time step. Comparing this net increase with the net
loss of the matching wait reward within the same time step can indicate whether holding
the passenger requests is a better option. Thus, this decomposition not only addresses the
sparse reward issue but also introduces a short-term signal to facilitate the learning process.
Therefore, the total pickup wait reward at tis
Rp(t) = Rp(S0, t) + Rp(S1, t).(9)
Summing up Eq. 7and Eq. 9, we obtain the total wait reward at time t:
Rw(t) = cmRm(t) + cpRp(t),(10)
where cmand cpare the perceptual cost per unit wait time. These weights can be set to
accommodate the asymmetrical perception of time values in different stages of the waiting
process (Guo et al.,2017). In general, cmcpas passengers are less patient in the face of
uncertain matching outcome (Xu et al.,2020a).
Qin et al. You can read the published version here Transp. Res. Part C
4.2. Learning optimal delayed matching policy through Actor-Critic methods
The intricacy of reinforcement learning in a realistic ride-hailing environment provokes
the need to appropriately adapt the general-purpose RL methods for improved robustness.
On account of this, we present two policy gradient-based RL methods along with three
aspects of meticulous adaptation.
First, outputting multi-binary actions. The actions taken in the conventional RL setting
are mostly discrete or continuous scalars, occasionally being continuous vectors, while the
action space in the ride-hailing environment, which is defined over a spatial network, is a
multi-binary matrix. To resolve this difficulty of representation, we propose a set of action
schemes A, as presented in Fig. 7, to map the one-dimensional action scheme ID (ranging
from 1 to 2|S|) chosen by the agent to the action matrix (ranging from a matrix of zeros to
ones) for feeding the environment. Similarly, we obtain the probability of choosing to match
in each grid by summing the learned policy probability π(A
A|s) of choosing each action scheme.
p(i) = X
where 1A
A(i)=1 is an indicator function that equals 1 if the action in grid i, i.e., A
A(i), is 1
(match), and equals 0 otherwise.
grid-based observations
input layer output layer π(a) action schemes
hidden layers
Figure 7: Structure of the policy neural network.
Second, increasing sample efficiency and reducing sample correlation. In the standard
RL setting, the agent’s experiences (observations, actions, and rewards) are usually utilized
once in value and policy updates. In the light of the large dimensionality of observation O
this one-time utilization can amount to a substantial waste of rare experiences and hinders
the rate of learning. We therefore adopt a technique called experience replay (Foerster et al.,
2017) to enable reuse of sampled experiences. Additional usefulness of experience replay is
the reduction of sample correlation. More specifically, the sampled observations are step-wise
Qin et al. You can read the published version here Transp. Res. Part C
correlated because each observation is updated upon the previous step’s carryover of supply
and demand. Experience replay blends the action distribution over the agent’s previous
observations, which can mitigate the divergence in parameters (Mnih et al.,2013).
Third, training agents with specific time limits. The interaction between an agent and
the ride-hailing environment has no terminal condition in standard RL literature. However,
the ride-hailing system’s matching requires setting time limits at the end of each sampling
period to smooth out the training processes. This coincides with the notion of episodes in
the experience sampling. It is worth noticing that, if the time limits are regarded as terminal
conditions, the RL algorithm will render an incorrect value function update. To distinguish
time limits from terminal conditions, we must perform a partial episode bootstrapping to
continue bootstrapping at the time limit. With this treatment, agents can learn the delayed
matching policy beyond the time limit (Pardo et al.,2018).
In the next subsections, we start by briefly introducing the general concept of the policy-
gradient methods. Then, we describe two policy gradient-based methods, Actor-Critic and
ACER, in which ACER is our solution to addressing those issues mentioned above.
4.2.1. Overview of policy gradient methods
Policy gradient methods are a family of RL methods that skip consulting the action
values and learn a parameterized policy function πθ(a|s) directly. In general, they have
many advantages over value-based methods, including stronger convergence guarantees and
the capacity of learning a stochastic policy (Sutton and Barto,2018). As discussed in the
section of designing observable state, using a stochastic policy can also obtain higher average
rewards than deterministic ones in POMDP.
Policy gradient methods define a performance measure J(θ), where θdenotes the param-
eterization of the policy in terms of the observable state, and improves it along the direction
of gradient ascent. The performance measure of a given policy is formally defined as
J(θ) = Vπθ(O
O(0)) ,
where the right-hand side value function denotes the expected return starting from state
O(0) = sby following sequences of actions taken from the policy πθ:= πθ(a|s).
The state value function Vπθ(O
O(0)) depends on both the policy and its changes on the
ride-hailing environment’s state distributions. However, the latter effect is unknown. As a
result, it is challenging to estimate J(θ). Fortunately, a solution in the form of the policy
gradient theorem has circumvented the derivative of the state distributions and provided an
analytic expression for J(θ),
qπ(s, a)πθ(a|s) = Eπ"X
O(t), a)πθ(a|O
where qπis the true action-value function for policy π, and η(s) is the on-policy distribution
of sunder policy π.
Following the gradient ascent method, the policy parameters are updated by
O(t), a)πθ(a|O
Qin et al. You can read the published version here Transp. Res. Part C
where ˆqis the learned approximation to qπ, and αis the learning rate.
Replacing ain Eq. (12) with A
A(t), we derive a vanilla policy gradient algorithm called
REINFORCE whose J(θ) can be computed by
J(θ)Eπ[G(t)ln πθ(A
O(t))] ,
where G(t) is the total matching wait reward from tonward.
4.2.2. Actor-Critic and ACER
Unlike REINFORCE and its variant methods, Actor-Critic estimates the total return
G(t) based on a bootstrapping, i.e., updating the value estimation of the current state from
the estimated value of its subsequent state,
G(t) := Vw(O
O(t)) R(t) + γVw(O
O(t+ 1)) .(13)
Bootstrapping the state value function Vw(·) (termed “Critic”) introduces bias and an
asymptotic dependence on the performance of the policy function (termed “Actor” ) estima-
tion. This is inversely beneficial for reducing the variance of estimated value function and
accelerates learning. To meet the needs of training in the time-limit ride-hailing environ-
ment, we adapt the Actor-Critic method, as summarized in Algorithm 1, to bootstrap at the
last observation in each episode by replacing the update on Line 6 with R+γVw(O0).
Algorithm 1: The Actor-Critic method
Input: πθ(a|s), Vw(s), step size αθ>0, αw>0,discount factor γ[0,1]
1Initialize policy parameter θRd, and state-value weights wRd0(e.g., to 0);
2for each episode do
3Initialize O(initial state of the episode), I1
4while Ois not terminal do
5Aπθ(·|O), take action A, observe O0, R
6δ[R+γVw(O0)] Vw(O)
8θθ+αθIδln πθ(A|O)
9IγI , O O0
10 end
11 end
ACER, standing for Actor-Critic with experience replay, is an upgrade to Actor-Critic
by enabling reuse of sampled experience (Wang et al.,2016). It integrates several recent
advances in RL to help stabilize the training process in the intricate ride-hailing environ-
ment. More specifically, ACER is a modification of Actor-Critic by adding a buffer to store
past experiences step by step for later sampling to batch update the value and policy neural
networks. Some other techniques, such as truncated importance sampling with bias correc-
tion, stochastic dueling networks and an efficient trust region policy optimization method,
are also added to enhance the learning performance. The proposed ACER method is also
adapted to the time-limit ride-hailing simulator by enabling bootstrapping at the last obser-
vations. Since the entire algorithm is rather intricate and out of the scope of this study, we
just present a general framework of ACER in Algorithm 2to give a broad picture of how it
Qin et al. You can read the published version here Transp. Res. Part C
Algorithm 2: Actor-Critic with experience replay (ACER)
Input: Dense NN Actor(a|s), Critic(s), and TargetCritic(s), discount factor γ[0,1], experience
replay batch size N, delayed update rate τ
1for each episode do
2Initialize O(initial state of the episode), buffer
3while Ois not terminal do
4AActor(·|O), take action A, observe O0, R from the ride-hailing simulator
5buffer.push(O, A, R, O0)// Collect experience
6// On-policy update
7Follow Actor-Critic update
10 // Off-policy update
11 if buffer.size > N then
12 O, A, R, O0= buffer.sample(N)// Sample experience to replay
13 δ[R+γTargetCritic(O0)] Critic(O)
14 CriticLoss δ2,ActorLoss δln Actor(A|O)
15 CriticLoss.minimize(),ActorLoss.minimize()
16 TargetCritic.w0= (1 τ)TargetCritic.w0+τCritic.w// Delayed update
17 end
18 end
4.2.3. Baseline: fixed matching time interval strategy
Industries currently implement a fixed matching time interval throughout an extended
period. More specifically, the matchmaker matches batched passenger requests every ntime
steps. A batch-learning version for the fixed strategy is to run the experiments repeatedly
to generate sufficient samples for each matching time interval candidate and choose the one
that yields the minimum mean wait costs. The chosen fixed matching time interval strategy
is a useful baseline, especially when the arrival of passenger requests is spatiotemporally
It is worth mentioning that this comparison is unfair for the RL-policy as we assume
no prior knowledge is available at the beginning of the learning process. On the contrary,
the baseline has found the optimal fixed matching time interval using the historical data. A
fixed matching time interval can offer a reasonable upper bound wait cost for reference in a
real-world scenario with non-homogeneous arrivals and complicated interdependence.
5. Result
In this section, we first delineate the experimental setup for testing the above-mentioned
policy gradient methods, then we present and compare the algorithmic performances of these
methods and the baseline strategy. Lastly, we analyze the learned policy and interpret how
it outperforms others.
5.1. Data description and experimental setup
We conduct a numerical experiment using a large-scale dataset from Shanghai Qiangsheng
Taxi Co. The one-week taxi dataset was collected in March 2011. It contains trip data,
sequential taxi trajectories, and operational status for nearly 9,000 taxis, with fields including
Qin et al. You can read the published version here Transp. Res. Part C
taxi ID, date, time, longitude, latitude, speed, bearing, and status identifier (1 for occupied/0
for idle).
0 5 10 15 20 0 5 10 15 20
Num. of requests per sec.
Hour of day
8:00 8:00
A1 B1
A2 B2
A3 B3
A1 B1
A2 B2
A3 B3
30 km
30 km
Figure 8: Spatial and temporal distributions of taxi supply and demand in downtown Shanghai, China.
Empirical evidence has shown that over 80% of taxi trips are concentrated in downtown
Shanghai outlined by the Outer Ring Road (Qin et al.,2017), which is selected as the area
of interest for our experiments. To balance the computational overhead incurred by the
dimension of actions, we partition the experiment area into six grids (3×2) as shown in Fig.
8. Experiments with finer grids are observed to increase the training time significantly with
marginal improvement on the overall performance. Three different partitioning granularities
of action are hereafter implemented and compared (i.e. 1 ×1, 1 ×2, and 3 ×2). In such a
partitioning, a desirable spatial variation of arrival patterns is well captured, with grid A2
producing high demand, grids A1, B1, and B3 producing low demand, and the remaining
grids producing medium-demand. The simulation initiates at 8:00 AM, which is the start of
the morning peak hours.
The ride-hailing environment is initialized with trip requests and taxi supply sampled
from the historical dataset as shown in Fig. 9. Then, the environment warms up for the
first 10 minutes, during which time all passengers and drivers are matched instantaneously.
Once the warm-up is complete, the learning agents begin interacting with the environment
through the simulator. The agents must match requests coming within the next 10 minutes,
which is considered an episode. After every episode, the environment is reset and re-warmed
up before the agents resume the learning process.
In terms of the algorithm implementation, we use a library called Stable Baselines (Hill
et al.,2018) to train the models. Additionally, we set unit wait cost cm= 4 and cp= 1 to
reflect a reasonably higher perceived value of the matching wait time over the pickup wait
time, and set average pickup speed v= 20km/h.
Qin et al. You can read the published version here Transp. Res. Part C
Taxi trip data
Trip OD between 7:50 and 8:10 AM
Taxi location at 7:50 AM
id, time, olat, olon, dlat, dlon
xxx,xxx, xxx, xxx, xxx, xxx
taxi_id, lat, lon, status
xxx, xxx, xxx, xxx
RL agents
Episode 1
Episode n
WARMUP (7:50~8:00)
Instantaneous matching MODEL TRAINING (8:00~8:10)
learned policy
Figure 9: Experimental setup of the ride-hailing environment and model training.
5.2. RL algorithm performance
5.2.1. Revisiting trade-offs in delayed matching
As a complement to the previous explanation of matching delay trade-off under stationary
arrivals, we now justify how this trade-off is established with real-world data. The wait time
curves under different driver occupancies are illustrated in Fig. 10(a), where the driver
occupancy is a parameter set to mimic a particular supply-demand ratio (100% corresponds
to the driver supply observed in the historical data). The result shows that the total wait
costs are constantly increasing under either a high (80%) or low driver occupancy3(50%)
(an instantaneous matching is preferred for this case). Under a moderate driver occupancy
(70%), the total wait cost indicates a U-shape curve.
Since the vehicle trajectory data is from a traditional street-hailing taxi market with
significant empty car mileage, feeding 100% of driver supply into the simulator corresponds
to a high-supply scenario. As a result, the search frictions are lower, and the required number
of drivers in the context of a ride-hailing platform should also be smaller. In light of this, the
finding here is consistent with the previous results under stationary arrivals – it is necessary
and beneficial to perform the delayed matching in the balanced supply-demand scenario.
This trade-off is more comprehensible after further separating the total wait costs into the
3Under a low driver occupancy (50%), some passengers experience a long wait beyond their tolerance
and choose to renege the queue. Reneging destabilizes the matching wait time in two opposite aspects: first,
those reneged passengers have experienced a longer matching wait than otherwise they had been matched
in time; second, the matching of the remaining passengers is expedited to some extent. Since the extent of
these impacts varies with matching time intervals, both the wait cost (Fig. 10(a)) and the matching wait
time (Fig. 10(b)) turn out to be zigzag.
Qin et al. You can read the published version here Transp. Res. Part C
(a) (b)
Matching wait time
Pickup wait time
driver occupancy
Fixed matching time interval
Relative changes
wait times
460 800
440 400
8 420 Q)
380 -125
100 120 0
100 120
Matching time interval (s) Matching time interval (s)
Figure 10: The relationship between the total wait costs and the matching time interval based on the fixed
matching time interval strategy under different driver occupancies: (a) total wait costs and (b) relative
changes to the matching and pickup wait times under one-second matching time interval. The lines stand
for the mean wait times from experiments and the shaded regions represent 95% confidence interval.
pickup and matching wait times, as seen in Fig. 10(b). Under different driver occupancies,
the relative increases of matching wait times (dotted curves) are almost at a fixed rate,
while in contrast, the corresponding decreases of the pickup wait times (solid curves) vary
drastically. Such a distinction suggests that the slope of the pickup wait times hugs the U-
shape of the matching delay curve. More specifically, this is because the pickup wait times
decrease much faster than the fixed increase rate of the matching wait times.
A key implication is that the desirable U-shape of matching delay trade-off merely appears
under a specific range of supply-demand ratios. The U-shape trade-off curve’s existence
condition, from a theoretical perspective, is worth further study; for example, future work
could model when an agent should apply instantaneous matching and when the agent should
otherwise follow the learned delayed matching policy.
5.2.2. Algorithmic performance
Following the discussion above, we find that sampling 70% of the existing taxi drivers is
a desirable value to prompt the matching delay trade-off. Using this driver-passenger ratio
as the instance for each RL algorithm, we repeat the model training for 1 million steps. The
training curves and their comparison with that of the fixed matching time interval strategy
are presented in Fig. 11.
In the baseline model, matching every 40 seconds is found to be optimal, and the served
passengers’ total wait cost is around 371.20 hours, as seen in Fig. 11(a). Note that the
Qin et al. You can read the published version here Transp. Res. Part C
Figure 11: Performance comparison between (a) the fixed matching time interval strategy and (b) policy
gradient algorithms: Actor-Critic with experience replay (ACER, main) and Actor-Critic (inset).
trade-off curve is by no means smooth with real-world data, while the change of wait costs
is subject to perturbation of matching time intervals. Learning a near-optimal adaptive
delayed matching policy is quite challenging to this end. The proposed algorithm ACER
under partitioning 3 ×2 reports the final converged wait cost, which is on a par with the
baseline, while ACERs under partitioning 1×2 and 1×1 fail to converge (Fig. 11(b)). These
results suggest that partitioning the area assists the agents in learning better policies. In
comparison, the original Actor-Critic algorithm only converges under the partitioning 1 ×2
to the wait cost around 380 hours, however its cost variance is greater than the result of the
ACER algorithm.
Table 1: Comparison of model performances
Algorithm Mean matching wait time Mean pickup wait time Mean total wait time
Instant. 3.01 675.71 678.72
Fixed int. 21.20 525.69 546.89
ACER (3 ×2) 26.95 513.22 540.17
Note: “Instant.” stands for instantaneous matching, “Fixed int.” for fixed matching time interval
matching, and ACER for Actor-Critic with delayed update and experience replay. Wait time unit:
s/trip request.
Table 1presents the wait time measures of the baseline strategies and the proposed ACER
algorithm (AC algorithm is excluded due to its mediocre performance). Compared with the
instantaneous matching strategy, the measures show that the learned ACER policy can
substantially reduce the mean pickup wait time. This policy can reduce total wait time by
20.41%, which is on a par with previous results in Ke et al. (2020) (292.60 221.88 : 24.17%
reduced). It is worth mentioning that such a comparison with prior work’s RL algorithms is
unjust, as the environment settings and data sources are different. However, it is plausible
to state that, as compared with the multi-agent model in Ke et al. (2020), our single-agent
Qin et al. You can read the published version here Transp. Res. Part C
model has a significantly smaller state space, but achieves a similar performance.
Mean match wait sec Mean pickup wait sec Mean total wait cost
08:00 14:00 19:00 08:00 14:00
Hour 19:00 08:00 14:00 19:00
Figure 12: Testing the learned policy. Baseline: fixed matching time interval strategy. Bar height indicates
mean value, error bar 95% confidence interval.
In addition, we test the performance of the learned policy for other hours and days in
Shanghai. We show that the learned policy is robust within and across days. Specifically, we
evaluate the performance of the ACER policy, which is learned with real-world data at 08:00
on a single weekday, in other weekday or weekend hours. The performance is compared with
the baseline strategy using fixed matching time interval. As Fig. 12 shows, total wait costs
yielded by the ACER policy are close to the baseline’s at all weekday 08:00 and weekend
14:00. This indicates that the learned policy generalizes well within and across days of those
periods. In the remaining periods, the ACER policy does not generalize well. However, this
can be reconciled by taking instantaneous matching because these periods have sufficient
supply and hence the instantaneous matching is more desirable. This case can be seen from
the near-zero matching wait time in the baseline. This can be an extension of our study to
introduce a meta-model that switches the platform’s matching policy between instantaneous
and delayed matching, mainly affected by the driver occupancy.
To summarize, the proposed two-stage ACER model has demonstrated its capacity to
reduce passenger wait costs and improve the quality of ride-hailing services.
5.3. Delayed matching policy interpretation
As the policy neural network is established on a high-dimensional state space, it is im-
possible to completely visualize the structure of the optimal policy in a low-dimensional
space. Instead, we try to interpret the learned policy by analyzing how the matching process
unfolds under a learned delayed matching policy. In other words, we replay episodic samples
by following the learned policy so as to reveal the decision-making step by step.
To be specific, we will track the matching probability in each grid along the episode to
reveal how the learned policy is responding to a certain environment state. Fig. 13 shows the
Qin et al. You can read the published version here Transp. Res. Part C
Matching probability
Matching probability
Matching probability
Matching probability
Matching probability
Matching probability
E[Δt] =
A1 B1
A2 B2
A3 B3
Matching probability along steps by following the learned optimal policy
1 - D
0.2 -
1 -
0.8 -
0.6 -
0.4 -
1 -
0.2 -
1 -
0.8 -f
0.6 -
. .
0.4 -
0.2 -:
1 - D
0.2 -
0 -
1 -
0.8 -
0.6 -
0 100 200 300
Step (s) 400 500 600 100 200 300
Step (s) 400 500
Figure 13: Episodic replay of a matching process under the learned policy for each grid.
Qin et al. You can read the published version here Transp. Res. Part C
grid-based matching probabilities and their corresponding environment state transitions. It
can be seen that the learned delayed matching policy appears to have different time-varying
matching probabilities (green curves) in different grids.
As for the three grids on the left (A1, A2 and A3), their matching probability curves look
alike on most parts and can be summarized in the following three stages. (a) Initial stage (0s
to 150s): they match passengers and drivers every 2s to 5s on average, since passengers in this
period arrive with a shortly increasing estimated rate λP(blue dotted curves), a mild delay
accumulating the arriving passengers is expected to shorten the pick up distance; (b) stable
stage (150s to 450s): they shorten the matching time interval and make an instantaneous
matching every 1s to 2s. Since arrival rates λPand λDat this stage are stabilized, while
the number of available drivers ND(magenta curves) is remarkably greater than the number
of waiting passengers NP(blue curves), the delayed matching becomes unnecessary and an
instantaneous matching is called for; (c) final stage (450s to 600s): A1 first transitions to 2s
to 5s and then returns to 1s to 2s, since its NDNPconsidering the lower level of both λP
and λD, an instantaneous matching is still desirable. By contrast, both A2 and A3 extend
their matching time interval beyond 10s, since their NDNPconsidering the higher level
of both λPand λD, a delayed matching is expected under such a condition.
As for the three grids on the right (B1, B2 and B3), their matching probability curves can
be summarized in two stages. (a) Initial stage (0s to 300s): Both B1 and B2 begin matching
with a matching time interval beyond 10s, intending to accumulate available drivers to
shorten the pickup distance, while B3 makes an instantaneous matching due to very few
passengers and driver expected to arrive into this grid; (b) final stage (300s to 600s): all
of the grids choose to make a matching every 1s to 2s, since at this stage their NPND
considering the lower level of both λPand λD. It is worth noting that compared with B2 and
B3, B1 takes an even shorter matching interval (almost 1s) due to its very stable observation
which lacks the benefit of delayed matching.
In general, the intuition behind the learned policy is consistent with the previous analysis
in Section 3.2. This implies that incorporating a queueing-based model can intrinsically
capture the delayed matching trade-off in the real-world ride-hailing platforms and adaptively
delay the matching to reduce passengers’ average wait costs. Integrating a structured model
into a model-free RL framework is promising in improving purely data-driven approaches.
6. Conclusion
This paper presents a family of policy gradient RL algorithms for the delayed matching
problem. Leveraging the sequential information collected by alternating the matching time
intervals when assigning trip requests to idle drivers, the proposed RL algorithm balances
the wait time penalties and improved matching efficiency. With prior information on the
stationary demand and supply distributions, we first characterize the matching delay trade-
off in a double-ended queue model. A multi-server queue model is then used to build a ride-
hailing simulator as a model-training environment, weighing the informativeness against
dimensionality. Searching for the optimal delayed matching policy in this environment is
formulated as a POMDP. We also propose a dedicated step-wise reward decomposition to
address the issue of the sparse reward. Finally, we unfold the learned policy into episodic
trajectories to garner insights.
Qin et al. You can read the published version here Transp. Res. Part C
We compare proposed RL algorithms’ performances with a baseline strategy using fixed
matching time interval and prior work based on real-world taxi data. Our numerical results
are two-fold. First, we identify the range of supply-to-demand ratio in which the learned
delayed matching policies outperform the instantaneous matching strategy. Second, the
learned delayed matching policy is both efficient and interpretable. Our results show that the
ACER algorithm can reduce the mean total wait time by amount on a par with the baseline.
More importantly, it can lower the wait time by 20.41% compared with the instantaneous
matching. The learned delayed matching has physical implications. A ride-hailing platform
can incorporate the structure of learned strategy as a lookup table to adaptively decide when
to use a delayed matching policy and how long these matching delays should be.
This paper leaves several extensions for future research. First, it is worth further inves-
tigation and development of an analytical model to quantify the U-shape trade-off curve at
a certain supply-and-demand ratio. To obtain this threshold, we must understand how the
first-stage matching time interval decision connects to the second-stage bipartite matching
rewards. Second, we can extend to a meta-model that switches the agent’s matching policy
between instantaneous and delayed matching and includes different factors such as passen-
gers’ utility into the objective functions. Finally, online RL is needed if implementing the
algorithm in the ride-hailing platform, which is challenging due to the platform’s stochas-
ticity. A model must be developed to reflect the platform’s real-time dynamics so that the
agents can learn effectively by interacting with it. However, the platform’s stochasticity may
complicate the agent’s target to learn, thereby resulting in an unstable RL algorithm with
no convergence guarantees.
The work described in this paper was partly supported by research grants from the
National Science Foundation (CMMI-1854684; CMMI-1904575) and DiDi Chuxing. The
first author (G. Qin) is grateful to the China Scholarship Council (CSC) for financially
supporting his visiting program at the University of Michigan (No. 201806260144). We also
thank Shanghai Qiangsheng Taxi Company for providing the taxi GPS dataset.
Al-Abbasi, A.O., Ghosh, A., Aggarwal, V., 2019. Deeppool: Distributed model-free algo-
rithm for ride-sharing using deep reinforcement learning. IEEE Transactions on Intelligent
Transportation Systems 20, 4714–4727.
Alonso-Mora, J., Samaranayake, S., Wallar, A., Frazzoli, E., Rus, D., 2017. On-demand high-
capacity ride-sharing via dynamic trip-vehicle assignment. Proceedings of the National
Academy of Sciences 114, 462–467.
Aouad, A., Sarita¸c, ¨
O., 2020. Dynamic stochastic matching under limited time, in: Proceed-
ings of the 21st ACM Conference on Economics and Computation, pp. 789–790.
Ashlagi, I., Burq, M., Dutta, C., Jaillet, P., Saberi, A., Sholley, C., 2018. Maximum weight
online matching with deadlines. arXiv preprint arXiv:1808.03526 .
Qin et al. You can read the published version here Transp. Res. Part C
Azar, Y., Ganesh, A., Ge, R., Panigrahi, D., 2017. Online service with delay, in: Proceedings
of the 49th Annual ACM SIGACT Symposium on Theory of Computing, pp. 551–563.
Cao, P., He, S., Huang, J., Liu, Y., 2020. To pool or not to pool: Queueing design for
large-scale service systems. Operations Research .
Foerster, J., Nardelli, N., Farquhar, G., Afouras, T., Torr, P.H., Kohli, P., Whiteson, S., 2017.
Stabilising experience replay for deep multi-agent reinforcement learning, in: International
conference on machine learning, PMLR. pp. 1146–1155.
Guo, S., Liu, Y., Xu, K., Chiu, D.M., 2017. Understanding passenger reaction to dynamic
prices in ride-on-demand service, in: 2017 IEEE International Conference on Pervasive
Computing and Communications Workshops (PerCom Workshops), IEEE. pp. 42–45.
Hare, J., 2019. Dealing with sparse rewards in reinforcement learning. arXiv preprint
arXiv:1910.09281 .
Hill, A., Raffin, A., Ernestus, M., Gleave, A., Kanervisto, A., Traore, R., Dhariwal, P.,
Hesse, C., Klimov, O., Nichol, A., Plappert, M., Radford, A., Schulman, J., Sidor, S., Wu,
Y., 2018. Stable baselines.
Jaakkola, T., Singh, S.P., Jordan, M.I., 1995. Reinforcement learning algorithm for par-
tially observable markov decision problems, in: Advances in neural information processing
systems, pp. 345–352.
Ke, J., Yang, H., Ye, J., 2020. Learning to delay in ride-sourcing systems: a multi-agent
deep reinforcement learning framework. IEEE Transactions on Knowledge and Data En-
gineering .
Kim, D.S., Smith, R.L., 1995. An exact aggregation/disaggregation algorithm for large scale
markov chains. Naval Research Logistics (NRL) 42, 1115–1128.
Li, M., Qin, Z., Jiao, Y., Yang, Y., Wang, J., Wang, C., Wu, G., Ye, J., 2019. Efficient
ridesharing order dispatching with mean field multi-agent reinforcement learning, in: The
World Wide Web Conference, pp. 983–994.
Lin, K., Zhao, R., Xu, Z., Zhou, J., 2018. Efficient large-scale fleet management via multi-
agent deep reinforcement learning, in: Proceedings of the 24th ACM SIGKDD Interna-
tional Conference on Knowledge Discovery & Data Mining, pp. 1774–1783.
Luo, Q., Huang, Z., Lam, H., 2019. Dynamic congestion pricing for ridesourcing traffic: a
simulation optimization approach, in: 2019 Winter Simulation Conference (WSC), IEEE.
pp. 2868–2869.
Mao, C., Liu, Y., Shen, Z.J.M., 2020. Dispatch of autonomous vehicles for taxi services: A
deep reinforcement learning approach. Transportation Research Part C: Emerging Tech-
nologies 115, 102626.
Qin et al. You can read the published version here Transp. Res. Part C
Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D., Riedmiller,
M., 2013. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602
Ozkan, E., Ward, A.R., 2020. Dynamic matching for real-time ride sharing. Stochastic
Systems 10, 29–70.
Pardo, F., Tavakoli, A., Levdik, V., Kormushev, P., 2018. Time limits in reinforcement
learning, in: International Conference on Machine Learning, pp. 4045–4054.
Qin, G., Li, T., Yu, B., Wang, Y., Huang, Z., Sun, J., 2017. Mining factors affecting
taxi drivers’ incomes using gps trajectories. Transportation Research Part C: Emerging
Technologies 79, 103–118.
Sutton, R.S., Barto, A.G., 2018. Reinforcement learning: An introduction. MIT press.
Wang, H., Yang, H., 2019. Ridesourcing systems: A framework and review. Transportation
Research Part B: Methodological 129, 122–155.
Wang, Z., Bapst, V., Heess, N., Mnih, V., Munos, R., Kavukcuoglu, K., de Freitas, N., 2016.
Sample efficient actor-critic with experience replay. arXiv preprint arXiv:1611.01224 .
Williamson, D.P., Shmoys, D.B., 2011. The design of approximation algorithms. Cambridge
university press.
Xu, Z., Li, Z., Guan, Q., Zhang, D., Li, Q., Nan, J., Liu, C., Bian, W., Ye, J., 2018.
Large-scale order dispatch in on-demand ride-hailing platforms: A learning and planning
approach, in: Proceedings of the 24th ACM SIGKDD International Conference on Knowl-
edge Discovery & Data Mining, pp. 905–913.
Xu, Z., Yin, Y., Chao, X., Zhu, H., Ye, J., 2020a. A generalized fluid model of ride-hailing
systems. Working Paper, University of Michigan .
Xu, Z., Yin, Y., Ye, J., 2020b. On the supply curve of ride-hailing systems. Transportation
Research Part B: Methodological 132, 29–43.
Yan, C., Zhu, H., Korolko, N., Woodard, D., 2020. Dynamic pricing and matching in ride-
hailing platforms. Naval Research Logistics (NRL) 67, 705–724.
Yang, H., Qin, X., Ke, J., Ye, J., 2020. Optimizing matching time interval and matching
radius in on-demand ride-sourcing markets. Transportation Research Part B: Method-
ological 131, 84–105.
Zha, L., Yin, Y., Yang, H., 2016. Economic analysis of ride-sourcing markets. Transportation
Research Part C: Emerging Technologies 71, 249–266.
Qin et al. You can read the published version here Transp. Res. Part C
Appendix A. Summary of notations
Table A.2: Notation list of variables and parameters
Notation Description
tA discrete time step in {0,1,· · · , T }
NP(t) Number of unmatched passenger requests at t
ND(t) Number of idle drivers at t
S(t) State of the ride-hailing environment at t
A(t) Matching decision; conducting a bipartite matching if A(t) = 1 and holding if A(t)=0
p(s0|s, a) Transition probability from sto s0after taking a
R(t) Reward received from the ride-hailing environment at t
r(s, a, s0) Wait reward at s0transitioned from safter taking a
π(a|s) Policy function, the probability of taking action aat s
Vπ(s) State value function, the expected return starting from sby following π
γDiscount factor, γ[0,1]
qπ(s, a) Action value function, the expected total reward starting from sby taking aand then following π
O(t) Partial observation at t
λP(t) Estimated arrival rate of passenger requests at t
λD(t) Estimated arrival rate of idle drivers at t
nP(t1:t2) Actual number of passenger requests arriving between t1and t2
nD(t1:t2) Actual number of idle drivers arriving between t1and t2
NM(t) Number of matched passenger-driver pairs at t
V(t) Expected total reward received from the ride-hailing environment at t
Rm(t) Reward regarding the matching wait cost at t
Rp(t) Reward regarding the pickup wait cost at t
Rw(t) Reward returned at t, a combination of Rm(t) and Rp(t)
τp(·,·) Total pickup wait cost function
cmUnit matching wait cost
cpUnit pickup wait cost
θParameters of the policy function or weights of the actor neural network
wParameters of the state value function or weights of the critic neural network
πθ(a|s) Parameterized policy function or actor neural network
Vw(s) Parameterized state value function or critic neural network
J(θ) Policy performance measure
αθLearning rate for the actor neural network
αwLearning rate for the critic neural network
G(t) Sampled total reward at t
η(s) Distribution of s
S0,S1,SSubarea 0, subarea 1, and the entire area
(x, y) Grid (x, y) in a partitioned area
o(x, y, t) observation of grid (x, y) at t
NP(x, y, t), ND(x, y, t) Number of unmatched passenger requests/idle drivers of grid (x, y) at t
λP(x, y, t), λD(x, y, t) Estimated arrival rate of passenger requests/idle drivers of grid (x, y) at t
RP(S, t) Reward regarding the pickup wait cost of subarea Sat t,S∈ {S0,S1}
ResearchGate has not been able to resolve any citations for this publication.
Full-text available
Online matching between idle drivers and waiting passengers is one of the most key components in a ride-sourcing system. It is naturally expected that a more effective bipartite matching can be implemented if the platform accumulates more idle drivers and waiting passengers in the matching pool. A specific passenger request can also benefit from a delayed matching since he/she may be matched with closer idle drivers after waiting for a few seconds. Motivated by the potential benefits of delayed matching, this paper establishes a two-stage framework which incorporates a combinatorial optimization and multi-agent deep reinforcement learning methods. The multi-agent reinforcement learning methods are used to dynamically determine the delayed time for each passenger request, while the combinatorial optimization conducts an optimal bipartite matching between idle drivers and waiting passengers in the matching pool. Four tailored reinforcement learning methods, delayed multi-agent deep Q learning (Delayed-M-DQN), delayed multi-agent actor-critic (Delayed-M-A2C), delayed multi-agent Proximal Policy Optimization (Delayed-M-PPO), and delayed multi-agent actor-critic with experience replay (Delayed-M-ACER), are developed. Through extensive empirical experiments with a well-designed simulator, we show that the proposed framework is able to remarkably improve system performances, by well balancing the trade-off among pick-up time, matching time, successful matching rate.
Full-text available
In a ride-sharing system, arriving customers must be matched with available drivers. These decisions affect the overall number of customers matched, because they impact whether future available drivers will be close to the locations of arriving customers. A common policy used in practice is the closest driver policy, which offers an arriving customer the closest driver. This is an attractive policy because it is simple and easy to implement. However, we expect that parameter-based policies can achieve better performance. We propose matching policies based on a continuous linear program (CLP) that accounts for (i) the differing arrival rates of customers and drivers in different areas of the city, (ii) how long customers are willing to wait for driver pickup, (iii) how long drivers are willing to wait for a customer, and (iv) the time-varying nature of all the aforementioned parameters. We prove asymptotic optimality of a forward-looking CLP-based policy in a large market regime and of a myopic linear program–based matching policy when drivers are fully utilized. When pricing affects customer and driver arrival rates and parameters are time homogeneous, we show that asymptotically optimal joint pricing and matching decisions lead to fully utilized drivers under mild conditions.
Full-text available
With the availability of the location information of drivers and passengers, ride-sourcing platforms can now provide increasingly efficient online matching compared with physical searching and meeting performed in the traditional taxi market. The matching time interval (the time interval over which waiting passengers and idle drivers are accumulated and then subjected to peer-to-peer matching) and matching radius (or maximum allowable pick-up distance, within which waiting passengers and idle drivers can be matched or paired) are two key control variables that a platform can employ to optimize system performance in an online matching system. By appropriately extending the matching time interval, the platform can accumulate large numbers of waiting (or unserved) passengers and idle drivers and thus match the two pools with a reduced expected pick-up distance. However, if the matching time interval is excessively long, certain passengers may become impatient and even abandon their requests. Meanwhile, a short matching radius can reduce the expected pick-up distance but may decrease the matching rate as well. Therefore, the matching time interval and matching radius should be optimized to enhance system efficiency in terms of passenger waiting time, vehicle utilization, and matching rate. This study proposes a model that delineates the online matching process in ride-sourcing markets. The model is then used to examine the impact of the matching time interval and matching radius on system performance and to jointly optimize the two variables under different levels of supply and demand. Numerical experiments are conducted to demonstrate how the proposed modeling and optimization approaches can improve the real-time matching of ride-sourcing platforms.
In this paper, we define and investigate a novel model-free deep reinforcement learning framework to solve the taxi dispatch problem. The framework can be used to redistribute vehicles when the travel demand and taxi supply is either spatially or temporally imbalanced in a transportation network. While previous works mostly focus on using model-based methods, the goal of this paper is to explore the policy-based deep reinforcement learning algorithm as a model-free method to optimize the rebalancing strategy. In particular, we propose an actor-critic algorithm with feed-forward neural networks as approximations of both policy and value functions, where the policy function provides the optimal dispatch strategy and the value function estimates the expected costs at each time stamp. Our numerical studies show that the algorithm converges to the theoretical upper bound with less than 4% optimality gap, whether the system dynamics are deterministic or stochastic. We also investigate the scenario where we consider user priority and fairness, and the results indicate that our learned policy is capable of producing a superior strategy that balances equity, cancellation, and level of service when user priority is considered.
There are two basic queue structures commonly adopted in service systems: the pooled structure where waiting customers are organized into a single queue served by a group of servers, and the dedicated structure where each server has her own queue. Although the pooled structure, known to minimize the servers' idle time, is widely used in large-scale service systems, this study reveals that the dedicated structure, along with the join-the-shortest-queue routing policy, could be more advantageous for improving certain performance measures, such as the probability of a customer's waiting time being within a delay target. The servers' additional idleness resulted from the dedicated structure will be negligible when the system scale is large. Using a fluid model substantiated by asymptotic analysis, we provide a performance comparison between the two structures for a moderately overloaded queueing system with customer abandonment. We intend to help service system designers answer the following question: To reach a specified service level target, which queue structure will be more cost-effective? Aside from structure design, our results are of practical value for performance analysis and staffing deployment.
With the rapid development and popularization of mobile and wireless communication technologies, ridesourcing companies have been able to leverage internet-based platforms to operate e-hailing services in many cities around the world. These companies connect passengers and drivers in real time and are disruptively changing the transportation industry. As pioneers in a general sharing economy context, ridesourcing shared transportation platforms consist of a typical two-sided market. On the demand side, passengers are sensitive to the price and quality of the service. On the supply side, drivers, as freelancers, make working decisions flexibly based on their income from the platform and many other factors. Diverse variables and factors in the system are strongly endogenous and interactively dependent. How to design and operate ridesourcing systems is vital—and challenging—for all stakeholders: passengers/users, drivers/service providers, platforms, policy makers, and the general public. In this paper, we propose a general framework to describe ridesourcing systems. This framework can aid understanding of the interactions between endogenous and exogenous variables, their changes in response to platforms’ operational strategies and decisions, multiple system objectives, and market equilibria in a dynamic manner. Under the proposed general framework, we summarize important research problems and the corresponding methodologies that have been and are being developed and implemented to address these problems. We conduct a comprehensive review of the literature on these problems in different areas from diverse perspectives, including (1) demand and pricing, (2) supply and incentives, (3) platform operations, and (4) competition, impacts, and regulations. The proposed framework and the review also suggest many avenues requiring future research.
The success of modern ride-sharing platforms crucially depends on the profit of the ride-sharing fleet operating companies, and how efficiently the resources are managed. Further, ride-sharing allows sharing costs and, hence, reduces the congestion and emission by making better use of vehicle capacities. In this paper, we develop a distributed model-free, DeepPool, that uses deep Q-network (DQN) techniques to learn optimal dispatch policies by interacting with the environment. Further, DeepPool efficiently incorporates travel demand statistics and deep learning models to manage dispatching vehicles for improved ride sharing services. Using real-world dataset of taxi trip records in New York, DeepPool performs better than other strategies, proposed in the literature, that do not consider ride sharing or do not dispatch the vehicles to regions where the future demand is anticipated. Finally, DeepPool can adapt rapidly to dynamic environments since it is implemented in a distributed manner in which each vehicle solves its own DQN individually without coordination.