You are on page 1of 2

Intelligent Traffic Light Scheduling for Traffic

Control
Rishabh Jain Ashish Sharma
Computer Science and Engineering Mathematics and Computing
Indian Institute of Technology, Ropar Indian Institute of Technology, Ropar
2019csb1286@iitrpr.ac.in 2019mcb1213@iitrpr.ac.in

Abstract—This is a study of various researches in traffic B. Mean Field Game Approach


light scheduling for reducing congestion in a road network. We
evaluate three researches on traffic light scheduling in this paper. Theophile Cabannes et. al [3] propose a Mean Field Game
approach for the N vehicle dynamic routing game. First
they model the traffic flow as a mesoscopic model, meaning
I. I NTRODUCTION individual vehicles are considered but their interactions are
considered with the remaining traffic flow as a whole. The
A. Motivation game is reformulated as a Mean Field Game. The paper
demonstrates the MFG approach efficiently approximates the
Traffic lights are road safety devices used to regulate traffic
N-player dynamic routing game. It is proved in both ap-
across road intersections. Often, traffic lights cause significant
proaches Nash Equilibrium(it is characterized by no change in
delays and congestion in modern cities. The growing number
optimal payoff value of a player irrespective of it changing its
of traffic lights necessitates a complex scheduling algorithm
strategy when all other player strategies are fixed) is achieved.
for both- the traffic lights and the cars, which is a hard
This approach was not taken further due to it not allowing
problem; and hence, automatic systems are indispensable
vehicle flow in FIFO(first in first out) order.
nowadays for optimally tackling this task.
There is little research on traffic routing algorithms which
C. MPLight Architecture
counter the issue of scalability, coordination and feasiblity of
acquiring data. Chacha Chen et. al [1] have suggested a scalable
Scalability here means the extensibility of the the solution reinforcement learning solution to traffic light scheduling. In
to several cities without major degradation in performance. this work, the authors assign observers to each intersection,
Coordination here means information sharing between differ- which communicate observations to a centralised RL agent
ent intersection signals, as if all signals work independently, with a Deep-Q network through parameter sharing. The goal
decision of one may negatively impact the other. A paper by of the RL agent is to minimize the pressure(Section III.B),
C.Chen et. al., MPLight [1], ensures all these requirements; or, make vehicle flow more uniform, and hence increase the
and is discussed in the section. throughput of vehicles. This will in turn minimise the traffic
congestion at every intersection. Maintaining independent
B. Problem Definition observers for all intersections increases the scalability of the
solution.
The problem definition is to develop a reinforcement learn-
ing based solution for intelligent traffic light scheduling in DQN Agent Description:
order to minimize vehicle wait times. The agent observes the pressure across the traffic segments
coming in and going out of the intersections (12 in total),
II. R ELATED W ORKS and the current phase (the traffic light states).
Actions: The agent (traffic signal controller at an intersection)
A. Clustering based Approach chooses its actions from the 8 possible phases, denoting what
V.Pattanaik et. al [2] Our initial aim was to develop an next state it should go to.
algorithm for vehicles, so that they can find optimal routes Reward: The reward for each agent (at each intersection say
in a dynamic and partially observable environment. A crucial i) is the negative of the Intersection Pressure:
part of our approach was the congestion prediction algorithm.
r = −P i (1)
For this we came across this paper. The algorithm uses K-
Means Clustering on the local surroundings of the vehicle, The DQN outputs the Q-value based on the Bellman Equation:
to find traffic clusters . To quantify different clusters, traffic
density is calculated with help of convex hull algorithm. Q(st , at ) = R(st , at ) + γmax.Q(st+1 , at+1 ) (2)
The coordination happening here can be seen in Fig.1, where
parameters of single network are shared among all agents.

Fig. 3. Signal Phases

IV. F UTURE W ORK


In the future, integration with the CityFlow simulation
platform has to be implemented. Along with that, possible
improvements in DQ Network will be explored.
Further, as per feedback, methods will be implemented to
Fig. 1. MPLight Architecture make way for Emergency Vehicles in intelligent traffic light
routing environment.
III. P RELIMINARIES R EFERENCES
Following are the definitions of a few terms that will be [1] C. Chen, H. Wei, N. Xu, G. Zheng, M. Yang, Y. Xiong,
frequently discussed in this paper. K. Xu, and Z. Li, “Toward a thousand lights: Decentralized
deep reinforcement learning for large-scale traffic signal control,”
A. Model of an intersection Proceedings of the AAAI Conference on Artificial Intelligence,
vol. 34, no. 04, pp. 3414–3421, Apr. 2020. [Online]. Available:
A road intersection is modelled as a way road intersection, https://ojs.aaai.org/index.php/AAAI/article/view/5744
with each road having two lanes, in opposite direction. A [2] V. Pattanaik, M. Singh, P. Gupta, and S. Singh, “Smart real-time traffic
congestion estimation and clustering technique for urban vehicular roads,”
traffic light is present at a lane which directs traffic into the in 2016 IEEE Region 10 Conference (TENCON), 2016, pp. 3420–3423.
intersection. Refer Figure 2. [3] T. Cabannes, M. Lauriere, J. Perolat, R. Marinier, S. Girgin, S. Perrin,
O. Pietquin, A. M. Bayen, E. Goubault, and R. Elie, “Solving n-player
dynamic routing games with congestion: a mean field approach,” 2021.
[Online]. Available: https://arxiv.org/abs/2110.11943

Fig. 2. Intersection

B. Pressure at an intersection
The definition of pressure, given by C. Chen et. al. [1] is
described as follows. Let x(l, m) denote the difference of the
number of vehicles entering into road l and exiting road m at
intersection i. The pressure of an intersection is defined as
X
Pi = x(l, m) (3)
∀(l,m)∈s

where s in the set of valid signal phases


C. Signal Phases
A traffic signal phase s is defined as a set of permissible
traffic movements. There are 8 possible phases for a 4 way
intersection. Refer Figure 3.

You might also like