Traffic light control in non stationary environments based on multi

Traffic Light Control in Non-stationary Environments based on Multi
Agent Q-learning
Monireh Abdoos , Nasser Mozayani and Ana L. C. Bazzan
Abstract— In many urban areas where traffic congestion does
not have the peak pattern, conventional traffic signal timing
methods does not result in an efficient control. One alternative
is to let traffic signal controllers learn how to adjust the lights
based on the traffic situation. However this creates a classical
non-stationary environment since each controller is adapting to
the changes caused by other controllers. In multi-agent learning
this is likely to be inefficient and computationally challenging,
i.e., the efficiency decreases with the increase in the number of
agents (controllers). In this paper, we model a relatively large
traffic network as a multi-agent system and use techniques from
multi-agent reinforcement learning. In particular, Q-learning
is employed, where the average queue length in approaching
links is used to estimate states. A parametric representation of
the action space has made the method extendable to different
types of intersection. The simulation results demonstrate that
the proposed Q-learning outperformed the fixed time method
under different traffic demands.
I. INTRODUCTION
Traffic control is a challenging issue when it comes to
application of computational techniques in the real world.
The development of efficient mechanisms for traffic light
control is necessary, as the number of vehicles in urban
network increases rapidly. The objective of the signal control
is to increase intersection capacity, to decrease delays, and
at the same time, to guarantee the safety of traffic actors.
Moreover, it can reduce fuel usage and reduce emissions.
Signal control is one of the areas involved in the overall
effort known as intelligent transportation systems (ITS). ITS
can be implemented by some techniques. In the present paper
we use multi-agent systems and machine learning to develop
a traffic light control mechanism.
For transportation systems, concepts of intelligent agents
may be used in different parts of the system such as traffic
lights [1], vehicles [2], and pedestrians [3], as well as to
model the behavior of traffic system to describe the norm
violation and critical situation detection [4].
In order to manage the increasing volume of traffic, traffic
lights control by artificial intelligence (AI) techniques are
becoming more and more important. Some methods handle
control of traffic signals by predefined rule-based system [5],
fuzzy rules [6], and centralized techniques [7].
Multi-agent Systems is a subfield of AI that aims to
provide principles for construction of complex systems in-
volving multiple agents. Multi-agent systems have gained
significant importance because of their ability to model and
solve complex real-world problems.
Multi-agent systems provide mechanisms for communi-
cation, cooperation, and coordination of the agents. In [8]
the individual traffic control is modeled by coordinated
intelligent agents. The agents coordinate among themselves
based on the information received from each other to control
the network. Synchronization of traffic signals has also been
studied as coordinated agents in [9]. Cooperative multi-agent
system has been proposed to control the signals according to
the prediction of traffic volume in neighboring intersections
[10].
Regarding use of machine learning methods, in [11]
collaborative reinforcement learning has been presented to
provide an adaptive traffic control based on traffic pattern
observed from vehicle location data. For an overview on
applications and challenges regarding multi-agent learning
in traffic signal control, the reader is refered to [12].
Traditional reinforcement learning research assumes that
the environment is stationary and its dynamics are always
fixed. This is not the case regarding a real traffic network as
a stationary environment since traffic flow patterns change
dynamically over the time.
In the present paper, we use a reinforcement learning
approach which is model-free, namely Q-learning (more on
this in Section II). Contrary to other works that use this
method, here Q-learning is used to control the traffic lights in
a large and non-regular (i.e. non grid) network. In summary,
a network with 50 junctions arranged in a non grid network
is considered.
As it will be seen in the next section, Q-learning does not
require a pre-specified model of the environment. Hence, it
can be used in dynamic and non-stationary environment. Of
course, in this case the mathematical guarantees of conver-
gence no longer hold. However, because the environment is
non-stationary, this is not even desirable as one wants the
method to re-adapt to the environment when this changes.
Here, each agent is responsible to control the traffic
lights in one junction using local information only. Another
important characteristic of the present work is that it handles
a large state and action space, especially because contrarily
to other works we do not make simplifications (e.g. only
considering two phases). Here a 4-phase signal timing is
considered for each intersection.
The rest of this paper is organized as follows. Reinforce-
ment learning and in particular Q-learning is discussed in
Section II. Section III presents some related works that have
used reinforcement learning to control traffic lights. The
proposed Q-learning method and the network configuration
are presented in Section IV. Network configuration and
experimental results are respectively given in Section V and
VI. Finally the paper is concluded in Section VII.
2011 14th International IEEE Conference on
Intelligent Transportation Systems
Washington, DC, USA. October 5-7, 2011
978-1-4577-2197-7/11/$26.00 ©2011 IEEE 1580

II. REINFORCEMENT LEARNING AND Q-LEARNING
Unlike supervised and unsupervised learning, reinforce-
ment learning learns an optimal policy by perceiving states
of the environment and receiving information from the
environment. Reinforcement learning algorithms optimize
environmental feedback by mapping percepts to actions.
These are best suited for domains in which an agent has
limited or no previous knowledge. The Q-Learning is one
particular reinforcement learning algorithm that is model-
free. Put simply, this means that an agent has no previous
model of how states and actions are related to rewards. The
Q-Learning algorithm was proposed by Watkins [13]. As
mentioned, it is essentially independent of existing domain
knowledge. This property of Q-Learning makes it a favorable
choice for unexplored environments.
Classically, Q-Learning is modeled by using a Markov
decision process formalism. This means that the agent is
in a state s, performs an action a, from which it gets a
scalar reward r from the environment. The environment then
changes to state s according to a given probability transition
matrix T. The goal of the agent is to choose actions maxi-
mizing discounted cumulative rewards over time. Discounted
means that short term rewards are weighted more heavily
than distant future rewards. The discount factor, γ = [0..1],
awards greater weight to future reinforcements if set closer to
1. In Q-learning, Q(s,a) is a state-action value representing
the expected total discounted return resulting from taking
action a in state s and continuing with the optimal policy
thereafter.
Q(s,a) ←− β(r +γV(s ))Q(s,a) (1)
V(s ) ←− max
a
Q(s ,a) (2)
In Equation 1, β is the learning rate parameter and V(s )
is given by Equation 2. The discount factor determines the
importance of future rewards. Q(s,a) measures the quality
value of the state-action pair and V(s ) gives the best Q value
for the actions in next state, s . Q-learning helps to decide
upon the next action based on the expected reward of the
action taken for a particular state. Successive Q values are
built increasing the confidence of the agent as more rewards
are acquired for an action. As an agent explores the state
space, its estimate of Q improves gradually and Watkins and
Dayan have shown that the Q-Learning algorithm converges
to an optimal decision policy for a finite Markov decision
process [14].
III. REINFORCEMENT LEARNING FOR TRAFFIC LIGHT
CONTROL
Reinforcement learning offers some potentially significant
advantages and has become a good solution to control
single junction as well as network of traffic signals, in non-
stationary and dynamic environments.
Some researchers have tried to use a model-based rein-
forcement learning method for traffic light control. In [15],
transition model has been proposed that estimates waiting
times of cars in different states. In [16] Silva et al have
presented a reinforcement learning with context detection
that creates a partial model of the environment on demand.
Later, the partial models are improved or new ones are con-
structed. Adaptive reinforcement learning has been presented
to control a model free traffic environment in [17], [18].
Adaptive control of traffic light has also been studied in [19],
which uses a function approximation method as a mapping
between the states and signal timing.
Some researchers have introduced a reinforcement learn-
ing method based on communication between the agents. A
collaborative reinforcement learning introduced in [11] has
tried to exploit local knowledge of neighboring agents in
order to learn signal timings.
In [20] an interaction model based on game theory has
been studied. They defined two types of agents for traffic
signal control, intersection agents and management agents,
and constructed models of two agents. Q-learning was used
to build the payoff values in the interaction model. In the
interaction model, the renewed Q-values in the distributed
reinforcement Q-learning was used to build the payoff values.
Therefore, interaction has taken on from the action selection
between two agents.
Balaji et al have presented a method in which agents
try to compute a new phase length based on both local
and communicated information [21]. Q values were shared
between agents to improve the local observations and create
a global view.
In [22], two types of agents have been used to control a
five intersection network. Four intersections connected to a
central intersection were labeled as outbound intersections.
Two types of agents, a central agent and an outbound agent,
were employed. An outbound agent that follows the longest
queue collaborates with a central agent by providing relative
traffic flow values as a local traffic statistic. The relative
traffic flow is defined as the total delay of vehicles in a lane
divided by the average delay at all lanes in the intersection.
The central agent learns a value function driven by its
local and neighbors traffic conditions. It incorporates relative
traffic flow of its neighbors as a part of its decision-making
process.
Some other methods have been using communicated infor-
mation in state and reward estimation [23], [24], optimization
through cooperation based on information fusion technology
in a multi-level hierarchical structure [25].
Several researches have tried to realize distributed control
that learn optimal signal control over a group of signal
by a hierarchical multi-agent system. In [26] a two level
hierarchical multi-agent system has been proposed. Local
traffic agents are concerned with the optimal performance of
their assigned intersection. A second layer has been added as
a coordinator that supervises the local agents. Bazzan et al
[27] have presented a hierarchical reinforcement learning for
networks. A number of groups of three traffic light agents
each is first composed. These groups are coordinated by a
supervisor agents that recommend a joint action.
1581

TABLE I
EXAMPLE OF STATE SPACE
State number Link order State number Link order
State 1 l1 ≥ l2 ≥ l3 ≥ l4 State 13 l2 ≥ l3 ≥ l1 ≥ l4
It is not common to find a reinforcement learning based
method that uses a large network for simulation without any
simplification. Most of the works mentioned here have used
a grid based network with a small number of states and
actions. As mentioned, in the present paper, a Q-learning
based method is used, which considers not only a large
number of states and actions, but also a non-grid based, non-
regular network.
IV. PROPOSED METHOD
In the Q-learning based method used here, the traffic
network is considered as a system composed of intelligent
agents, where each controls an intersection.
As state estimation, agents use the average queue length
in approaching links in a fixed cycle. They then select an
action and receive a reward. The state space is discretized
by ranking the approaches according to the statistic gathered
in the last cycle.
The number of states is equal to the number of permutation
of the approaches, i.e., k! for an agent with k approaching
links. To see why this is is so, consider the following
example. There are 24 states for each junction, since each
agent has four approaching links. The state space can be
seen in Table I. If li represents the ith approaching link for
an intersection, l1 ≥ l2 ≥ l4 ≥ l3 shows a state in which the
queue length of approaching link l1 is the longest and l3 is
the shortest. If two approaching links have the same queue
length, then they are ranked according to their order in the
signal plan.
The actions relate to split of green time, i.e., the action
space contains different phase splits of the cycle time. Phase
split refers to the division of the cycle time into a sequence
of green signals for each group of approaching links. We
assumed that the cycle time, δ, is fixed and all junctions use
the same value.
It should be noted that for a fixed time controller, all
available phases should at least appear once in a cycle.
Each phase should have a minimum green time so that a
stopped vehicle that receives a green signal has enough time
to cross the intersection. The cycle length is divided to a
fixed minimum green time, and extension time that can be
assigned to different phases.
Fig. 1. Phase split parameters used for action definition
Assume that there is nph phases and minimum green time
assigned for each of them is the same and equal to tmin.
Moreover, there are nex extensions with fixed length of time,
hex. The actions determine the assignment of nex extensions
to different phases.
For example, assuming a cycle length of 50 seconds,
a signal plan containing 3 phases and 10 second as the
minimum green time, 3 × 10 seconds is assigned to the
phases and different actions determine the assignment of the
remaining 20 seconds to three phases.
The action space is defined by < nph,tmin,nex,hex >,where:
nph: number of phases tmin: minimum green time for
phases (seconds) nex: number of time extensions hex: length
of each extension (seconds) such that the effective cycle
length δ is given as in Equation 3.
δ = hex ×nex +nph ×tmin (3)
The possible green time assignment can be formulated as
follows:
nph
∑
i=1
xi = nex, xi ∈ 0,1,...,α, 1 ≤ α ≤ nph (4)
The maximum number of extension is controlled by α.
The action space can be reduced by choosing the small value
for α. The solutions determine the portion of the cycle time
that is assigned to each phase by:
di = tmin +xi ×hex (5)
In equation 5, di is the green time assigned to phase i.
Figure 1 shows the phase split parameter used for the
definition of the actions. For example, d1, the green time
assigned to the first phase, includes tmin as minimum green
time and a number of extension displayed by x1.
The reward is inversely proportional to the average length
of the queues in the approaching links, normalized to remain
between 0 and 1.
V. NETWORK CONFIGURATION
The traffic network used in the experiments is shown in
Figure 2. It contains 50 junctions and more than 100 links.
A four-way intersection is the most common intersection in
real world, and therefore it has been used in our approach.
The number of lanes is three for all approaches. During
the simulation, new vehicles are generated by uniform dis-
tribution over 30 input sections. Time intervals between two
consecutive vehicle arrivals at input sections are sampled
from a uniform distribution. The network is surrounded by
1582

Fig. 2. The network used for the simulation
Fig. 3. An example for action corresponds to [x1,x2,x3,x4] = [0,2,1,1].
Two seconds is assigned for all-red interval between two signals
20 centroids. Centroids are points that vehicles enter or exit
the network through them.
In this configuration, the number of phases is four and the
minimum green time for each phase is set to 13 seconds.
Moreover, four 10-second extension intervals are used for
signal timing, i.e. nex = 4 and hex = 10. The effective
cycle time is computed according to Equation 3 with δ =
92seconds.
The purpose of all-red interval is to provide a safe tran-
sition between two conflicting traffic signal phases. Two
seconds is assigned for all-red interval at the end of each
signal phase as it has been shown in Figure 3. Since there are
four phases, 8 seconds are assigned for all-red time interval.
In this case, the total cycle time will be 100seconds for a
complete signal plan.
The parameters used in the Q-learning based method are
shown in Table II. There are 19 possible solutions for
Equation 4 if we assume α = 2. For example, [x1,x2,x3,x4] =
[0,2,1,1] is a solution for Equation 4. The corresponding
action is (13,33,23,23) as shown in Figure 3.
The action set is as follows:
TABLE II
THE PARAMETERS USED IN THIS PAPER
Parameter Description Value
δ effective cycle length 92
nph number of phases 4
tmin minimum green time for each phase 13
nex number of extension interval 4
hex length of extended time 10
α maximum number of extension for each phase 2
|S| number of states 24
|A| number of actions 19
TABLE III
NETWORK CONFIGURATION PARAMETERS
Properties Value
number of intersections 50
number of links 224
average length of links 847.44m
number of lanes per links 3
maximum speed 50km/h
number of input/output centroid 20
arrival distribution Uniform
simulation duration 10hour
traffic demand 1 18000 veh/hour
traffic demand 2 27000 veh/hour
A = {a1(33,33,13,13),a2(33,13,23,23),...,a19(23,23,23,23)}
(6)
Moreover, we use ε-greedy as a mechanism to select an
action in Q-learning with ε fixed at value of 0.9. This means
that the best action is selected for a proportion 1 − ε of
the trials, and a random action is selected (with uniform
probability) for a proportion ε.
VI. EXPERIMENTAL RESULTS
The proposed method has been tested on a large network
that contains 50 intersections and 112 two-way links using
the Aimsun traffic simulator 1. The system configuration
parameters are shown in Table III.
We performed some experiments using Aimsun in order
to compare the proposed method with fixed time method
that assigns equal green time to each signal group. The
experiment aimed to determine the simulated average delay
time of the network for different traffic demands. Among
different traffic demands we choose two values as medium
and high traffic congestion (traffic demands 1 and 2 in
Table III).
In these experiments, the average delay time is used for
performance evaluation. The results show that the proposed
approach has reduced the average delay time in comparison
with the fixed time method. Figure 4 shows the result over
a traffic demand with 18000 vehicles per hour (demand 1).
The performance of the methods over a traffic demand
with 27000 vehicles per hour (demand 2) can be seen in
Figure 5. The average value of the delay is shown in Table
IV.
1http://www.aimsun.com
1583

Fig. 4. Comparison between fixed time method and the proposed Q-learning based approach for traffic demand 1
TABLE IV
AVERAGE DELAY OF THE PROPOSED QL AND FIXED TIME METHOD
Fixed time Proposed QL
Traffic demand 1 47.977 42.362
Traffic demand 2 113.034 63.687
For the fixed timing, performance begins to drop rapidly
when nearly all of the vehicles in some junctions stop. This
happens after six hours after the begin of the simulation, for
traffic demand 2, as seen in Figure 5. At the end of the
simulation, all vehicles stop in the network and there is no
vehicle leaving network. Under these conditions, the average
delay over all simulated time makes no sense. Therefore, the
value reported in Table IV for traffic demand 2 has been
computed over the period of the first 6 hours.
Regarding the proposed method, it is possible to see that
it could manage to efficiently control the traffic lights in the
network under relatively high traffic demand in a better way
than the fixed timing method.
VII. CONCLUSIONS
In this work we have shown that Q-learning is a promising
approach to control traffic light in a non-stationary environ-
ment. A real traffic network is a non-stationary environment
because traffic flow patterns are dynamically changed over
the time. The proposed method is based on local statistical
information gathered in one learning step and tries to learn
the best action in different situations. Average queue length
in approaching links is used as the statistical information.
The proposed method contains a relatively large number of
states and actions that can be used in a large network.
The simulation results demonstrated that the proposed Q-
learning method outperformed the fixed time method under
different traffic demands.
Parametric representation of the action space enables us to
define different situations by means of using different values
of the parameters. Since the action space is defined by a
number of parameters, it can be extended to a larger space
or smaller one by changing the parameters. As a result, it
can be applied in real traffic networks with a non predictable
traffic demand. Moreover, it can be used in different types of
intersections with a different number of approaching links.
Future work includes applying the method to other net-
works, extension of the action space with different parame-
ters to find the proper parameters for a given traffic demand,
and examination of impact of different parameters values on
learning performance.
VIII. ACKNOWLEDGEMENTS
A. Bazzan and M. Abdoos are partially supported by
CNPq. The present research was supported by Research
Institute for Information and Communication Technology -
ITRC (Tehran, Iran). The first author gratefully acknowledge
the support of this institute.
REFERENCES
[1] Z. Liu, “A survey of intelligence methods in urban traffic signal
control,” International Journal of Computer Science and Network
Security, vol. 7, no. 7, pp. 105–112, 2007.
[2] J. Adler, G. Setapathy, V. Manikonda, B. Bowles, and B. V.J., “A multi-
agent approach to cooperative traffic management and route guidance,”
Transportation Research Part B, vol. 39, pp. 297–318, 2005.
[3] K. Teknomo, “Application of microscopic pedestrian simulation
model,” Transportation Research Part F, vol. 9, pp. 15–27, 2006.
[4] A. Doniec, R. Mandiau, S. Piechowiak, and S. Espié, “A behavioral
multi-agent model for road traffic simulation,” Engineering Applica-
tions of Artificial Intelligence, vol. 21, no. 8, pp. 1443–1454, 2008.
[5] V. Hirankitti, J. Krohkaew, and C. Hogger, “A multi-agent approach
for intelligent traffic-light control,” in Proc. of the World Congress on
Engineering, vol. 1. London, U.K.: Morgan Kaufmann, 2007.
[6] I. Kosonen, “Multi-agent fuzzy signal control based on real-time
simulation,” vol. 11, no. 5, pp. 389–403, 2003.
[7] P. Balaji, P. Sachdeva, D. Srinivasan, and T. C.K., “Multi-agent system
based urban traffic management.” Singapore: IEEE, 2007, pp. 1740–
1747.
[8] R. T. v. Katwijk, B. De Schutter, and J. Hellendoorn, “Multi-agent
coordinationof traffic control instruments,” in Proc. of the International
Conference on Infrastructure Systems 2008: Building Networks for a
Brighter Future, vol. 1, Rotterdam, The Netherlands, November 2008.
1584

Fig. 5. Comparison between fixed time method and the proposed Q-learning based approach for traffic demand 2
[9] A. L. C. Bazzan, “A distributed approach for coordination of
traffic signal agents,” Autonomous Agents and Multiagent Systems,
vol. 10, no. 1, pp. 131–164, March 2005. [Online]. Available:
www.inf.ufrgs.br/∼bazzan/downloads/jaamas10 2 131 164.pdf.gz
[10] F. Daneshfar, F. Akhlaghian, and F. Mansoori, “Adaptive and coopera-
tive multi-agent fuzzy system architecture,” in Proc. 14th International
CSI Computer Conference. IEEE, 2009, pp. 30–34.
[11] A. Salkham, R. Cunningham, A. Garg, and V. Cahill, “A collaborative
reinforcement learning approach to urban traffic control optimization,”
in Proc. of International Conference on Web Intelligence and Intelli-
gent Agent Technology. IEEE, 2008, pp. 560–566.
[12] A. L. C. Bazzan, “Opportunities for multiagent systems and multiagent
reinforcement learning in traffic control,” Autonomous Agents and
Multiagent Systems, vol. 18, no. 3, pp. 342–375, June 2009. [Online].
Available: http://www.springerlink.com/content/j1j0817117r8j18r/
[13] C. Watkins, “Learning from delayed rewards,” Ph.D. dissertation,
University of Cambridge, 1989.
[14] C. J. C. H. Watkins and P. Dayan, “Q-learning,” Machine Learning,
vol. 8, no. 3, pp. 279–292, 1992.
[15] M. Weiring, “Multi-agent reinforcement learning for traffic light
control,” in Proc. of the Seventh International Conference on Machine
Learning, 2000, pp. 1151–1158.
[16] B. C. d. Silva, E. W. Basso, A. L. C. Bazzan, and P. M.
Engel, “Improving reinforcement learning with context detection,” in
Proceedings of the 5th International Joint Conference On Autonomous
Agents And Multiagent Systems, AAMAS, 2006. Hakodate, Japan:
New York, ACM Press, May 2006, pp. 811–812. [Online]. Available:
www.inf.ufrgs.br/maslab/pergamus/pubs/Silva+2006.pdf
[17] K. Wen, S. Qu, and Y. Zhang, “A stochastic adaptive control model for
isolated intersections,” in Proc. 2007 IEEE International Conference
on Robotics and Biomimetics. Sanya, China: IEEE, 2008, pp. 2256–
2260.
[18] Y. Dai, D. Zhao, and J. a. Yi, “A comparative study of urban traffic
signal control with reinforcement learning and adaptive dynamic
programming,” in Proc. of International Joint Conference on Neuarl
Networks, Barcelona, 2010, pp. 1–7.
[19] L. Prashanth and S. Bhatnagar, “Reinforcement learning with func-
tion approximation for traffic signal control,” IEEE Transaction on
Intelligent Transportation Systems, pp. 1–10, 2010.
[20] X. Xinhai and X. Lunhui, “Traffic signal control agent interaction
model based on game theory and reinforcement learning,” in Proc.
2009 International Forum on Computer Science-Technology and Ap-
plications. IEEE, 2009, pp. 164–168.
[21] P. Balaji, X. German, and D. Srinivasan, “Urban traffic signal control
using reinforcement learning agents,” IET Intelligent Transportation
Systems, vol. 4, no. 3, pp. 177–188, 2010.
[22] I. Arel, C. Liu, T. Urbanik, and A. Kohls, “Reinforcement learning-
based multi-agent system for network traffic signal control,” IET
Intelligent Transport Systems, vol. 4, no. 2, pp. 128–135, 2010.
[23] J. Medina, A. Hajbabaie, and R. Benekohal, “Arterial traffic control
using reinforcement learning agents and information from adjacent
intersections in the state and reward structure,” in Intelligent Trans-
portation Systems (ITSC), 2010 13th International IEEE Conference
on, sept. 2010, pp. 525–530.
[24] W. Yaping and Z. Zheng, “A method of reinforcement learning
based automatic traffic signal control,” in Proc. Third International
Conference on Measuring Technology and Mechatronics Automation.
IEEE, 2011, pp. 119–122.
[25] Z. Yang, X. Chen, Y. Tang, and J. Sun, “Intelligent cooperation control
of urban traffic networks,” in Proc. Fourth International Conference
on Machine Learning and Cybernetics. Guangzhou, China: IEEE,
2005, pp. 1482–1486.
[26] J. France and A. A. Ghorbani, “A multiagent system for optimizing
urban traffic,” in Proceedings of the IEEE/WIC International Confer-
ence on Intelligent Agent Technology. Washington, DC, USA: IEEE
Computer Society, 2003, pp. 411–414.
[27] A. L. C. Bazzan, D. d. Oliveira, and B. C. d. Silva, “Learning in groups
of traffic signals,” Engineering Applications of Artificial Intelligence,
vol. 23, pp. 560–568, 2010.
1585

Traffic light control in non stationary environments based on multi

More Related Content

What's hot (19)

Viewers also liked (10)

Similar to Traffic light control in non stationary environments based on multi (20)

Recently uploaded (20)

Traffic light control in non stationary environments based on multi