Velocity-Assisted Predictive Mobility and
Location-Aware Routing Protocols for
Mobile Ad Hoc Networks
Kai-Ten Feng, Member, IEEE, Chung-Hsien Hsu, Student Member, IEEE, and Tse-En Lu
Abstract—In recent years, many location-aware routing proto-cols have been proposed for mobile ad hoc networks. The effi-ciency of the routing protocols can be improved by considering the location information of the mobile nodes (MNs). However, the mobility characteristics of the MNs have not been taken into account in most of the related work. In this paper, the proposed velocity-aided routing (VAR) algorithm determines its packet-forwarding scheme based on the relative velocity between the intended forwarding node and the destination node. The routing performance can further be improved by the proposed predic-tive mobility and location-aware routing (PMLAR) algorithm, which incorporates the predictive moving behaviors of MNs in protocol design. The region for packet forwarding is determined by predicting the future trajectory of the destination node. The routing performance can be effectively enhanced by adopting both the proactive maintenance and the VAR mechanisms within the proposed PMLAR scheme. Simulation results show that the PMLAR protocol associated with its derivative schemes outper-forms other routing protocols under different network topologies. Index Terms—Location-based ad hoc routing protocol, Mobile Ad hoc NETwork (MANET), mobility model, prediction mechanism.
I. INTRODUCTION
A
MOBILE Ad hoc NETwork (MANET) consists of wire-less mobile nodes (MNs) that cooperatively communicate with each other without the existence of a fixed network in-frastructure. Depending on different geographical topologies, the MNs are dynamically located and continuously changing their locations. The fast-changing characteristics of MANETs make it difficult to discover routes between MNs. It becomes important to design efficient and reliable routing protocols to maintain, discover, and organize the routes in MANETs. Recent interest in the design of ad hoc routing algorithms include appli-cations for the military, intervehicle communication, personal communication services, and sensor networks.Manuscript received January 9, 2006; revised August 10, 2006 and February 25, 2007. This work was supported in part by the National Science Council under Grant 92-2218-E-009-018, the MOE ATU Program 95W803C, and the MediaTek Research Center, National Chiao Tung University. The review of this paper was coordinated by Prof. A. Ganz.
K.-T. Feng and C.-H. Hsu are with the Department of Communication Engineering, National Chiao Tung University, Hsinchu 300, Taiwan, R.O.C. (e-mail: [email protected]; [email protected]).
T.-E. Lu is with Alpha Networks Inc., Science-based Industrial Park, Hsinchu 300, Taiwan, R.O.C. (e-mail: [email protected]).
Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TVT.2007.901897
A number of ad hoc routing protocols have been developed for MANETs. The topology-based routing protocols can be categorized into proactive (such as destination sequence dis-tance vector [1] and wireless routing protocol (WRP) [2]) and reactive algorithms (such as ad hoc on-demand distance vector routing [3], dynamic source routing (DSR) [4], temporally-ordered routing algorithm (TORA) [5], associativity-based routing (ABR) [6], and signal stability based adaptive routing [7]). The proactive routing protocols periodically maintain ta-bles at each MN. The routing tata-bles record persistent and up-to-date information within the changing network topologies. The proactive algorithms offer reliable routing information between the MNs, while the overhead for maintaining the routing tables can be rapidly increased as the expansion of the MN’s num-bers and mobility within the network. The reactive algorithms initiate route discovery processes based on the request from the source node. The routing tables are only maintained within the requested routes from the source to the destination nodes. The on-demand characteristics of the reactive algorithms gen-erate less routing overhead compared with the proactive routing protocols. However, additional delay has been incurred by the route discovery processes within the reactive algorithms.
The hybrid routing approaches (such as zone routing protocol (ZRP) [8], [9], clusterhead gateway switch routing (CGSR) [10], [11], and core extraction distributed ad hoc routing algorithm (CEDAR) [12]) compromise the benefits between proactive and reactive algorithms. ZRP utilizes the proactive-based approach within its predetermined local area, while the reactive algorithm is adopted outside the area. Cluster heads (as in CGSR) or cores (as in CEDAR) are selected from the MNs to serve as gateways for information distribution within the network. The hybrid approaches can be utilized to adopt themselves by considering the tradeoffs between the communication reliability and the routing overhead. The per-formance comparison between the various types of ad hoc routing protocols has been conducted in several studies, as in [13]–[16].
There are increasing interests in the design of location-based
ad hoc routing algorithms [17]–[25]. With the prosperity of
mo-bile devices equipped with positioning systems (such as global positioning systems [26]), it becomes feasible to utilize the location information from mobile devices in routing protocol design (such as distance routing effect algorithm for mobility (DREAM) [20], location-aided routing (LAR) [21], geograph-ical routing algorithm (GRA) [22], greedy perimeter stateless routing (GPSR) [23], and grid location service (GLS) [24],
[25]). The DREAM protocol proactively distributes the location information of each MN within the network. The data packets are delivered from the source node to the destination node based on the built-in location database within each MN. On the other hand, the LAR protocol is a reactive location-based algorithm. The route discovery and packet-forwarding region is restricted within a predetermined area based on the knowledge of the MN’s location. The routing overhead can therefore be reduced by the directional flooding scheme proposed in the protocol.
As indicated in the survey on position-based routing algo-rithms [27], the location services and the packet-forwarding strategies are the two major components in on-demand routing protocol design. The primary purpose of location services is to provide the position information of the destination node to the source node for its packet delivery. Based on the in-formation, the position-aware routing protocols can determine their route discovery and packet-forwarding strategies. Most of the position-aware routing protocols utilize global flood-ing [20]–[22] as their location services, which results in ex-cessive amounts of control packet overhead. The algorithm alleviates the problem by providing location updates within the local area of the network. A small set of MNs serves as the location server that offers location information for their neighborhood MNs. The appropriate selection of the location servers influences the performance of the location updates within the GLS scheme. A comparative study between the location-based ad hoc routing protocols can be obtained as in [28] and [29].
On the other hand, different types of MN mobility models have been studied in the previous research works [30]–[32]. It has been observed and demonstrated that the mobility models utilized in the ad hoc network simulations greatly influence the effectiveness of the routing algorithms. The commonly used random waypoint mobility (RWM) model (as was uti-lized in [13] and [14]) has been found to be insufficient in most of the realistic scenarios [33]. Different types of mobility models have been proposed to emulate the motion behaviors of MNs. It is suggested that the designed ad hoc routing protocols should be evaluated under the various types of MN mobility models to satisfy the requirements for realistic circumstances.
However, the mobility patterns of the MNs have not been taken into consideration in most of the location-based routing protocol design. Since the velocity and heading angle of the MN are obtainable from positioning systems, it is practicable to incorporate the information in the design of ad hoc routing protocols. The proposed velocity-aided routing (VAR) algo-rithm determines its packet-forwarding scheme by calculating the relative velocity between the potential forwarding nodes and the destination node. This scheme forwards the data pack-ets via those intermediate nodes that are faster approaching the destination node. The Gauss–Markov mobility (GMM) model [34], [35] and the constant speed mobility (CSM) model are utilized in the design of the VAR algorithm to calculate the MN’s speed and moving angle. The benefit of using the GMM model is that it can be utilized to adaptively emulate possible moving behavior with certain levels of linear and Brownian motions. It will be shown in this paper that the
pro-posed VAR algorithm is especially feasible for topologies with confined shapes and dynamic characteristics (e.g., highways or city streets). It will also be illustrated in the simulation results that the VAR algorithm based on the GMM model provides better routing performance compared with that based on the CSM model.
Moreover, since the mobility of the MNs is diverse under dif-ferent moving scenarios, it will be beneficial to incorporate the MN’s predicted movement in the design of routing protocols. The proposed predictive mobility and location-aware routing (PMLAR) protocol determines its packet-forwarding scheme by predicting the future position of the destination node. The prediction mechanism defines the packet-forwarding region by adapting its adjustable parameters based on the previous moving behavior of the destination node. The GMM model is utilized as the prediction model in the design of the PMLAR algorithm, where its tunable parameters are adjusted with on-line adaptation. There are three additional enhanced schemes associated with the PMLAR algorithm. The PMLAR-V scheme incorporates the VAR algorithm to consider the relative velocities between the MNs within the route discovery process. The PMLAR-LV method further encompasses the local repair-ing mechanism within the PMLAR-V scheme. Furthermore, the PMLAR-PV scheme enhances the routing performance by exploiting both the VAR algorithm and the mechanism for proactive maintenance. It will be shown in this paper that the proposed PMLAR algorithm with its derivative schemes is fea-sible for different types of network topologies. Three mobility models, i.e., the highway mobility (HM) model, the RWM model, and the RWM model with indirect route (RWM-IR), are implemented in the simulations to provide different types of simulation scenarios. The effectiveness of the proposed VAR, PMLAR, PMLAR-V, PMLAR-LV, and PMLAR-PV algorithms is evaluated and compared with other existing routing protocols via simulations.
The remainder of this paper is organized as follows: Section II-A and B review and reformulate the existing DSR and LAR protocols. The two predictive mobility models, i.e., the GMM and CSM models, are presented in Section III-A. Section III-B explains the proposed VAR algorithm based on the two predictive mobility models. The proposed PMLAR protocol and its derivative schemes are presented in Section IV. The mobility models for simulation purposes, including a fea-sible HM model, are developed and introduced in Section V-A. The simulation parameters are defined in Section V-B. The parameter estimation for both RWM and HM models is con-ducted in Section V-C, while the performance evaluation of the proposed routing protocols is demonstrated Section V-D. Section VI draws the conclusions.
II. PRELIMINARIES
In this section, two on-demand routing protocols, i.e., the DSR and LAR algorithms, are reviewed. These two protocols provide the baseline route discovery and data transmission mechanisms for the proposed PMLAR protocol. Both protocols will be analyzed and compared with the PMLAR algorithms in the simulations. The route discovery and packet-forwarding
strategies of the DSR and LAR protocols are summarized as follows.
A. DSR Protocol
The DSR protocol [4] is a source-initiative on-demand rout-ing algorithm. The source node S intends to send data packets to the destination node D via some of the forwarding nodes Ni.
S will first check its route cache to verify if there are existing
routes to the destination node D. If there is no such route in the cache, S will start a route discovery process by broadcasting a route request (RREQ) packet. The addresses of S and D are both written in the route record in the RREQ packet header. Upon receiving the RREQ packet, the intermediate node Ni
will rebroadcast the packet if the route to D is not available.
Ni will also add its own address to the route record in the
RREQ packet. Until the destination node D is found in the
route discovery process, a route reply (RREP ) packet, which contains the complete route information from S to D, is sent back to the source node S. The source node S will utilize the constructed route to initiate data transmission to D.
B. LAR Protocol
The LAR protocol is also an on-demand routing algorithm [21]. With the assistance of the MN’s position information, the algorithm restricts the packet flooding area in the route discovery process instead of spreading out to the whole region as in the DSR protocol. The position information of the desti-nation node D is assumed available by the source node S in the LAR protocol. One of the commonly used LAR algorithms, which is called the LAR-Box, defines a rectangular region for packet forwarding named the Request Zone. The circle, which is called the Expected Zone in the LAR-Box algorithm, is centered at the destination position PD(tp), where tprepresents
a previous time instant that the location information of D was acquired by S. The radius of the Expected Zone (Rl)
is determined by S at the current time instant tc, which is
defined as Rl= Vave∆t. It is noted that Vave represents the
average speed of D within the time duration ∆t = (tc− tp).
Within the limited region (i.e., the Request Zone determined by
S), the LAR-Box algorithm follows similar procedures as the
DSR protocol to choose the appropriate intermediate nodes Ni
for packet forwarding. The efficiency for route search can be increased with the constrained area of packet flooding.
III. VAR PROTOCOL
It will be shown in this section that the knowledge of the destination node’s predicted motion is required in the design of the VAR algorithm. The algorithm is designed by considering the following two types of motion prediction models.
A. Mobility Models
In this subsection, two mobility models, i.e., the GMM model and the CSM model, are presented. These two mobility models will be utilized as two separate cases to predict the
MN’s movement in the design of the proposed VAR routing algorithm.
1) GMM Model: The GMM model [34] is adapted in this
paper to represent the motion of each MN. The moving direc-tion αk(with respect to the positive x-axis) and the speed Vkof
each MN at a discrete time instant tkcan be formulated as [35]
αk = γ1αk−1+ (1− γ1) ¯α + (1− γ2 1)Xαk−1 (1) Vk = γ2Vk−1+ (1− γ2) ¯V + (1− γ2 2)XVk−1 (2)
where ¯α and ¯V represent the asymptotic mean of the moving
direction and the speed as tk → ∞; Xαk−1 and XVk−1 are
the zero-mean Gaussian-distributed random variables; and γ1
and γ2 are tunable parameters that represent different levels
of randomness as 0≤ γi≤ 1 for i = 1 and 2. The two
ex-treme cases correspond to the linear motion (as γi= 1) and
the Brownian motion (as γi= 0). The benefit of using the
GMM model for the MN’s movement is that it preserves certain levels of 1) motion randomness and 2) memories from previous time steps as the parameter γi varies. A more realistic motion
trajectory can be obtained with the GMM model compared with the Random Walk Mobility Model [36], which results in sharp turns and sudden stops. With (1) and (2), the position P (xk, yk)
of the MN at time instant tkbecomes
xk = xk−1+ Vk−1δt cos αk−1 (3)
yk = yk−1+ Vk−1δt sin αk−1 (4)
where δt is the sampling interval between the time instant tk
and its previous time instant tk−1. It is assumed that the position
information of an MN, including its position P (xi, yi), velocity
Vi, and heading angle αi, is obtained at time instant ti via
its positioning system. P (xj, yj) is the predicted position of
the MN at time instant tj as tj≥ ti. The traveling distance
along the x and y directions (i.e., ∆xi,j and ∆yi,j) of the MN
within the time duration ∆ti,j= (tj− ti) can be obtained by
summing along both directions from (3) and (4) as
∆xi,j= j k=i+1 Vkδt cos αk (5) ∆yi,j= j k=i+1 Vkδt sin αk. (6)
It is noted that αk and Vk represent the moving direction and
speed of D at the time instant tk, which can be obtained from
(1) and (2).
2) CSM Model: In the previous subsection, the GMM
model is utilized to obtain the predicted information (including the speed Vjand the moving direction αj) of an MN based on
its previous motion states. However, in some cases, it is possible to assume that the MN remains a constant speed (Vo) along the
same moving direction (αo) within the time interval ∆ti,j=
(tj− ti), i.e., V (t) = Vo and α(t) = αo for t∈ [ti, tj]. The
traveling distance of the MN along the x and y directions within the time duration ∆ti,j can therefore be simplified from (5)
Fig. 1. Schematic of the VAR scheme using the GMM model.
and (6) as
∆xi,j= Vo∆ti,jcos αo (7)
∆yi,j= Vo∆ti,jsin αo. (8)
The constant speed assumption in the CSM model is applica-ble while the MN moves in a constant speed (along with the same direction) manner. This scenario is possible to happen in constrained topologies with similar traffic flows, including city streets and highway topologies. With the constant speed assumption, the CSM model provides a less-expensive compu-tation cost compared with the GMM model. The CSM model is treated as a special case for simple motion prediction. The routing performance of the VAR algorithm using the CSM model will be compared with that using the GMM model in the simulations.
B. VAR Algorithm
In this section, the proposed VAR algorithm is presented. The proposed algorithm determines the feasible MNs for packet for-warding based on the relative velocity between the forfor-warding node Ni and the destination node D. The VAR algorithm is
designed by predicting the motion of D using either the GMM model or the CSM model, as shown in the next two subsections.
1) VAR Using the GMM Model (VAR-GMM): Fig. 1 shows
the schematic of the VAR algorithm using the GMM model. The source node S initiates a route discovery process to the destination node D via some of the intermediate nodes Ni.
After beaconing within the neighborhood of S, an RREQ packet is sent to an intermediate node Niat the time instant tc.
The location information of Ni, including its position PNi(tc),
velocity VNi(tc), and heading angle αNi(tc), is obtained from
its positioning system at the current time instant tc. On the
other hand, the location information of D [i.e., PD(tp), VD(tp),
and αD(tp)] was obtained by S at a previous time instant tp
and was forwarded to Ni via the RREQ packet. By adopting
the GMM model, the current location information of D [i.e.,
PD(tc), VD(tc), and αD(tc)] can therefore be calculated from
the previous time instant tpusing (1)–(4).
The main concept of the VAR algorithm is to compare the ve-locity information between the intermediate node Ni with that
of D while the RREQ packet has arrived in Ni at the current
time tc. The proposed VAR algorithm utilizes the following
two criterions to determine if the intermediate node Ni is
suitable as the forwarding node for packet delivery. The first criterion [as in (9)] of the VAR algorithm indicates that the potential forwarding node Ni should move toward the
desti-nation node D along their connecting line (i.e., the ith direction as shown in Fig. 1), while the second criterion [as in (10)] is used to limit the relative speed between Ni and D along their
perpendicular direction j, i.e.,
∆Vi > [VNi(tc) + VD(tc)] δ1,i+ δ2,i (9) |∆Vj| < [VNi(tc) + VD(tc)] δ1,j+ δ2,j (10) where ∆Vi= VNi(tc) cos (θ(tc)− αNi(tc)) − VD(tc) cos (θ(tc)− αD(tp)) (11) ∆Vj = VNi(tc) sin (θ(tc)− αNi(tc)) − VD(tc) sin (θ(tc)− αD(tp)) . (12)
As shown in Fig. 1, θ(tc) represents the angle between D and
Niat tc, which can be calculated as
θ(tc) = tan−1
[yD(tp)− yNi(tc)] + (∆yp,c)D
[xD(tp)− xNi(tc)] + (∆xp,c)D
(13) where the location information (xD(tp), yD(tp)) of D is
ob-tained from the positioning system at the previous time instant
tp, and the location (xNi(tc), yNi(tc)) of Niis acquired at the
current time tc. Both (∆yp,c)Dand (∆xp,c)Dcan be obtained
by using (5) and (6). It is also noted that δ1,i, δ2,i≥ 0 and δ1,j, δ2,j > 0 in (9) and (10) are the tuning parameters for the
VAR criterions. δ1,iand δ1,j are utilized to represent the
speed-dependent tuning coefficients. With the VAR criterions, the amount of potential forwarding nodes within the constrained flooding area will decrease.
2) VAR Using the CSM Model (VAR-CSM): As described
in Section III-A2, it is feasible to assume the MN to pertain to a constant speed during the time interval (t∈ [tp, tc]) in
consideration. The VAR criterions [i.e., (9)–(12)] proposed in the previous subsection are still applicable to be adopted in the CSM model. The only difference is that a simplified version of
θs(tc) is obtained as shown in Fig. 2, i.e.,
θs(tc) = tan−1
[yD(tp)− yNi(tc)] + VD(tp)∆tp,csin αD(tp)
[xD(tp)− xNi(tc)] + VD(tp)∆tp,ccos αD(tp)
(14) where ∆tp,c= (tc− tp). It can be seen that the computation
complexity of θs(tc) is much less than the θ(tc) obtained from
(13). In the simulation section, the performance comparison between the VAR-GMM and the VAR-CSM schemes will be conducted under different simulation mobility models.
Fig. 2. Schematic of the VAR scheme using the CSM model.
IV. PMLAR PROTOCOL
The proposed PMLAR protocol will be explained in this section. The first subsection describes the three phases of the PMLAR protocol in detail. The second subsection presents the PMLAR algorithm with the assistance of the VAR scheme, i.e., the PMLAR-V protocol. The PMLAR-LV scheme is ad-dressed in the third subsection, which incorporates a local repair mechanism within the PMLAR-V method. The fourth subsection incorporates a proactive feedback information from
D to S within the PMLAR-V algorithm to improve the routing
efficiency, i.e., the PMLAR-PV protocol. The following list of parameters denotes the timing variables that will be utilized in the PMLAR protocols:
t(n)l,s time instant while sending the LREQ packet for the
nth route request;
t(n)l,f time instant while receiving the LREP packet for the
nth route request;
t(n)r,s time instant while sending the RREQ packet for the
nth route request;
t(n)r,f time instant while receiving the RREP packet for the
nth route request;
t(n)p,s starting time instant while delivering data packets for
the nth route request;
t(n)p,f ending time instant while delivering data packets for the nth route request;
t(n)e time instant while receiving the RERR packet for the
nth route request;
t(n)m time instant while receiving the proactive maintenance
packet for the nth route request.
A. PMLAR Protocol
Fig. 3 shows the schematic of the network topology with the proposed PMLAR protocol. The PMLAR algorithm is designed for S to predict the current and future location of D to achieve efficient data transmission. An enhanced location service is also considered in the proposed algorithm to address the potential indirect route (IR) between S and D. The three phases of
Fig. 3. Schematic of the network topology with the PMLAR protocol.
the PMLAR algorithm (i.e., the location service, the route discovery and packet forwarding, and the route repair phases) are stated as follows.
1) Location Service: The source node S intends to transmit
data packets to the destination node D. However, S has no knowledge about the position information of D at the beginning. It is required for S to activate a location service to obtain the position information of D. As shown in Fig. 4, the location service request packet (LREQ(n)p ) initiated by
S at time t(n)l,s is transmitted to its neighboring MNs using the flooding mechanism, i.e.,
LREQ(n)p =AdrS, AdrD (15)
where the superscript (n) denotes the nth route request, and the subscript p represents the PMLAR protocol. The parameters
AdrSand AdrDdenote the addresses of nodes S and D. After
rebroadcasting by the intermediate node Ni, D is informed that
a position request has been initiated by S. It is assumed that the position information of D is updated at the time instant
t(n)l , where t(n)l ={t ∈ |0 < t ≤ t(n)l,f}. The up-to-date infor-mation of D, including the time stamp t(n)l , position PD(t(n)l ),
velocity PD(t(n)l ), moving angle αD(t(n)l ), and parameters
γiD(t(n)l ) (for i = 1 and 2), is sent back to S via the reverse
route. Moreover, the position information of the forwarding nodes Ni, i.e., PNi(t(n)l ) = [PN1(t
(n)
l ), . . . , PNn(t(n)l )], is also
transmitted back to S. The positions of the intermediate nodes
PNi(t(n)l ) will be utilized in the route discovery phase to
deter-mine if an indirect region for packet forwarding is exadeter-mined. The location service reply packet (LREPp(n)) is received by S
at time t(n)l,f while the location information of D is available, i.e.,
LREPp(n)= AdrS, AdrD, PD t(n)l , VD t(n)l αD t(n)l , γiD t(n)l , PNi t(n)l . (16)
Fig. 4. Timing diagram of the source node S using the PMLAR protocol and its derivative schemes.
It is noted that the two tunable parameters of D, γiD(t(n)l ) are
computed and acquired at time instant t(n)l . As discussed in Section III-A, γ1Dand γ2D represent the level of randomness
for the corresponding moving angle (αk) and speed (Vk) of D.
To facilitate solving these two parameters, both (1) and (2) can be combined and rewritten into the following format:
y k = Hkγk+ vk (17) where y k = αk Vk , Hk= αk−1 α¯ 0 0 0 0 Vk−1 V¯ γ k = γ1k 1− γ1k γ2k 1− γ2k , vk= (1− γ2 1)Xαk−1 (1− γ2 2)XVk−1 .
The state vector ykcontains the moving direction αkand speed
Vkof D at time instant tk. Hkis the design matrix for parameter
estimation, while γk represents the state vector for the time-varying parameters γ1and γ2of D at time tk. vk denotes the
system noises that are scaled from the random variables Xαk−1
and XVk−1. The two parameters γ1and γ2can be estimated at
each time instant tk (i.e., ˆγk) by solving the state (17) using
the recursive least square (RLS) estimation [38], [39] as ˆ γk+1= ˆγk− Kk+1 Hk+1ˆγk− yk+1 (18) with Kk+1= Pk+1HTk+1 Pk+1= 1 λ Pk− PkHTk+1Hk+1Pk λ + Hk+1PkHTk+1 .
It is noted that the tunable parameter λ determines the convergence rate of the RLS method. The appropriate values of
λ under different mobility environments will be determined in
the simulation section. As the parameters ˆγ1Dand ˆγ2Dare
ob-tainable with online adaptation, the prediction mechanism can forecast the future position of D based on the state information.
2) Route Discovery and Packet Forwarding: S can start the
processes of route discovery and forwarding of data packets after executing the location service. The prediction mechanism utilized in the proposed PMLAR algorithm is for S to predict the trajectory of D from its previous location update. As shown in Fig. 3, the trajectory T (i.e., the sold path) shows the actual
moving path of D, while T (i.e., the dotted path) indicates the predictive path of D starting from the time instant t(n)l . As mentioned in the previous subsection, t(n)l corresponds to the time instant that the location information of D is acquired from the positioning system. The current and future positions of D can be obtained at different predictive time steps as ∆Tl(n)=
κ∆t(n)l = κ(t(n)r,s − t(n)l ), where t
(n)
r,s corresponds to the current
time instant for initiating a route request packet (RREQ(n)p ).
As κ = 1, the current position of D is obtained, while the future position of D is predicted if κ > 1. The prediction mechanism of the PMLAR protocol defines a Predicted Zone based on the predicted trajectory of D, as illustrated in Fig. 3. The Predicted
Zone is defined as a circular region centered at PD(t(n)l ) with
its radius R(n)p defined as
Rp(n)= max
0<i≤ζ
[(∆xl,i)D]2+ [(∆yl,i)D]2
1 2
(19) where (∆xl,i)Dand (∆yl,i)Dcan be obtained from (5) and (6),
and ζ = ∆Tl(n)/δt represents the total number of predicting
steps. The distances from PD(t(n)l ) to the position of the ith
predicting step PD(t(n)i ) are calculated, and the maximum
value is selected as the radius of the Predicted Zone. As shown in Fig. 3, the predicting mechanism starts at the time instant
t(n)l , and the predicting trajectory by using the GMM model is shown in dotted line with empty circles. The number of predicting steps ζ is equal to 7 in this case. The radius Rp(n)
represents the predicted radius that happens at the fifth steps, as in Fig. 3. Similar to the Expected Zone as in the LAR-Box algorithm, the Predicted Zone defines a circle that predicts the potential movement of the destination node D. However, the determination of the Predicted Zone is based on the predictive moving behavior of D, while the LAR-Box algorithm assumes a constant moving speed of D along the time interval in consideration.
As the Predicted Zone is computed by S, either the direct routing (DR) or the IR type will be determined. The parameter
µ [as in (20), shown at the bottom of the next page] is utilized to
examine the types of routing that is going to be applied, where the position information of D and the intermediate nodes are acquired from LREPp(n), as in (16). The parameter µ is utilized
to represent the ratio of the routing distances to the connecting line between S and D. Two cases are considered as follows.
1) If µ≤ µt(where µtis a prespecified threshold), the DR
type of routing is utilized. Similar to the LAR-Box algo-rithm, a rectangular region, which encloses S’s position
Fig. 5. Schematic of the PMLAR algorithm with the IR scheme.
and the circle of radius R(n)p , is determined, as shown in
Fig. 3. The Request Zone (RZp−DR(n) ) can be obtained as
RZp(n)−DR= ΓDR PS t(n)r,s , PD t(n)l , R(n)p (21) where the function ΓDR(·) defines the rectangular area
for the DR type of routing.
2) If µ > µt, the IR type of routing will be adopted. A
dif-ferent type of Request Zone for the IR type is illustrated in Fig. 5. The positions of the intermediate nodes [i.e.,
PNi(t(n)l )] are utilized as the waypoints for the confined
region. The Request Zone (RZp−IR(n) ) for the IR type of routing [defined by the function ΓIR(·)] is obtained as
RZp(n)−IR= ΓIR PS t(n)r,s , PD t(n)l , R(n)p , PNi t(n)l . (22) After the process of location services, S initiates a route discovery process to D via some of the intermediate nodes Ni.
An RREQ(n)p packet will be sent out by S to an intermediate
node Niat the time instant t(n)r,s as
RREQ(n)p = AdrS, AdrD, RZp−DR(n) , for µ≤ µt
AdrS, AdrD, RZp(n)−IR
, for µ > µt
(23) where the types of Request Zone (i.e., either RZp−DR(n) or
RZp(n)−IR) selected in the RREQ(n)p packet will depend on the
magnitude of the parameter µ. It is utilized for the intermediate
MN to determine if it itself is located within the confined region. After the route to D is established with the appropri-ate selection of the forwarding MNs, an RREPp(n) packet is
returned to S at time t(n)r,f as
RREPp(n)=
AdrS,{AdrNi} , AdrD, PD
t(n)r VD t(n)r , αD t(n)r , γiD t(n)r (24) where{AdrNi} denotes the addresses for the set of
intermedi-ate MNs that are selected as the forwarding nodes within the route. The additional information of D [including PD(t(n)r ),
VD(t(n)r ), αD(t(n)r ), and γiD(t(n)r )] is also delivered back to
S via the RREPp(n) packet, where t(n)r ={t ∈ |t(n)l < t≤
t(n)r,f} represents the time instant that the location information
is updated at D. Thereafter, the source node S will utilize the route obtained from the RREPp(n)packet to transmit data
packets to D between the time instants t(n)p,s and t(n)p,f.
3) Route Repair: Due to the dynamic changes of network
topologies, route maintenance and repair should take place during data transmission from S to D. The proposed PMLAR algorithm provides a repairing mechanism for broken transmis-sion links.
Similar to the route error repairing capability as in the DSR protocol [4], the source node S receives a route error (RERR(n)p ) packet from one of the forwarding MNs if there
exists a broken transmission link in the originally defined route. As shown in Fig. 4, the link is broken at the time instant
t(n)e , where t(n)e ={t ∈ |t(n)p,s < t < t(n)p,f} happens between
the duration of data packet transmission. A new route discovery process [i.e., the (n + 1)th route request] will be initiated by
S. The source node S will define a new Request Zone (i.e.,
either RZp−DR(n+1) or RZp−IR(n+1)) as the confined region to select the appropriate MNs for packet forwarding. A new set of route request/reply process will be adopted, which is followed by the remaining data packet transmission.
B. PMLAR Protocol With VAR Scheme (PMLAR-V)
The prediction mechanism, as described in the PMLAR protocol, defines the Predicted Zone that predicts the potential future position of the destination node D. All of the inter-mediate nodes within the Request Zone (as shown in Fig. 3) are considered as potential forwarding nodes to transmit date packets, e.g., node Ni, Nj, or Nk. However, it is possible to
further impose additional constraints on the selection of the in-termediate nodes. Since the velocity information of the MNs is available from most of the positioning systems, it is reasonable to incorporate the MN’s velocity information as an additional
µ = PN1 t(n)l − PS t(n)r,s +nk=1−1PNk+1 t(n)l − PNk t(n)l +PD t(n)l − PNn t(n)l PD t(n)l − PS t(n)r,s (20)
criterion to determine if the MN is a feasible node for packet forwarding from S to D. A new route request (RREQ(n)p−v) packet will be sent out by S at time t(n)r,s to initiate the route
discovery process. By combining the location information of D [including PD(t(n)l ), VD(t(n)l ), and αD(t(n)l )] that was obtained
from the LREPp(n) packet [as in (16)], the original route
request (RREQ(n)p ) packet [as in (23)] is extended as that in
(25), shown at the bottom of the page, where the subscript
p− v denotes the PMLAR-V algorithm. After receiving the RREQ(n)p−vpacket from S, Niwill perform the VAR algorithm
[based on the criterions in (9) and (10)] to determine if it itself is a suitable node for packet forwarding. If the VAR criterions are satisfied, Niwill record itself within the routing information
on the RREQ(n)p−v packet header and rebroadcast the packet to
Nj. The destination node D is informed that a request for data
transmission is initiated by S after receiving the RREQ(n)p−v
packet. D will send out the route reply (RREPp−v(n)) packet to
S via the reverse route, which is recorded on the RREQ(n)p−v
packet header. The RREPp−v(n) is returned to S at time t(n)r,f as RREPp(n)−v =RREPp(n), where RREPp(n)is defined as
in (24). After completing the route discovery process, the data packets can be delivered from S to D via the designated route. As shown in Fig. 3, the route is selected to comply with both the prediction mechanism and the VAR algorithm.
It is noted that in the original VAR algorithm, the route re-quest packet is flooded to all regions for route discovery, where they seek for intermediate nodes that satisfy the criterions as in (9) and (10). With the limited searching area within the Request Zone in the PMLAR-V algorithm, the inefficiency occurring from the all-region flooding can be improved. The selection of the potential forwarding nodes can therefore be confined within the predefined region.
C. PMLAR Protocol With VAR and Local Repair Schemes (PMLAR-LV)
In this subsection, a local repair mechanism is adopted within the PMLAR-V scheme to improve the overhead induced by the original repairing process. Assuming that the route for packet delivery has been constructed as
RT ={S, . . . , Ni, Nj, . . . , D}, the link between nodes Niand
Nj is considered broken due to the dynamic movements of
the MNs. In the original route repair scheme (as stated in Section IV-A3), the source node S will receive a route error (RRER(n)p ) packet from node Ni, and a new route will be
dis-covered and initiated by node S for further data transmission. In the local repair scheme, node Niis considered as the source
node after it receives the route error message. Node Ni will
start to initiate the route request packet targeting to the
destina-tion node, which now becomes node Nj. Therefore, the route
request/reply process will be locally conducted between nodes
Ni and Nj instead of the end-to-end route repair (between
nodes S and D), as in the original scheme. The performance of the PMLAR-LV scheme by adopting the local repair mecha-nism will be evaluated in the simulation section.
D. PMLAR Protocol With VAR and Proactive Maintenance Schemes (PMLAR-PV)
This subsection introduces a mechanism for proactive route maintenance before the linkage within a route is likely to break. The prediction mechanism of the destination node D monitors its own position to ensure that it will stay within the
Predicted Zone, i.e., the circle shown in Fig. 3. The prediction
mechanism periodically observes the updated position of D from the positioning system to verify if D is still located within the boundary of the Predicted Zone. If D moves close to (or goes out of) the boundary at time instant t(n)m , the
mecha-nism for proactive maintenance within D will be initiated. As shown in Fig. 3, PD(t(n)m ) indicates the position of D that
is close to the boundary of the circle with radius R(n)p . The
updated information of D, including t(n)m , PD(t(n)m ), VD(t(n)m ),
αD(t(n)m ), and γiD(t(n)m ), is sent back to S using the reversed
data transmission route (i.e., within the P M(n) packet, as
shown in Fig. 4). S will initiate a new route discovery process with the RREQ(n+1)p−pv packet, which is similar to (25) but with
the updated information of D at time t(n)m . The subscript of
RREQ(n+1)p−pv denotes a PMLAR protocol with the VAR and proactive maintenance schemes.
It is noted that both RZp(n+1)−DR and RZp(n+1)−IR, which are defined in the RREQ(n+1)p−pv packet, rely on the newly generated
Predicted Zone, i.e., the dotted circle with radius Rp(n+1), as
shown in Fig. 3. The proposed PMLAR-PV protocol provides the mechanism for D to update its current information to S if it moves toward the boundary of the Predicted Zone. The proac-tive update scheme can effecproac-tively increase the communication robustness within the delivering route. The performance of the proposed algorithms (including the VAR, PMLAR, PMLAR-V, PMLAR-LV, and PMLAR-PV schemes) will be simulated and evaluated in the next section.
V. PERFORMANCEEVALUATION
A. Mobility Model
Different types of MN’s mobility affect the performance of the designed ad hoc routing algorithms. It is therefore important to construct feasible mobility models that emulate the realistic moving environment for simulation purposes. The RWM model
RREQ(n)p−v= AdrS, AdrD, RZp−DR(n) , PD t(n)l , VD t(n)l , αD t(n)l , for µ≤ µt
AdrS, AdrD, RZp(n)−IR, PD
t(n)l , VD t(n)l , αD t(n)l , for µ > µt (25)
is widely used to evaluate the performance of ad hoc routing protocols [13], [14], [36]. Each MN moves toward a randomly selected destination node with a chosen speed. The MN pauses for a preselected timeout period and resumes its movement again. The MN’s speed and timeout period are tunable parame-ters to simulate different moving environments.
In this paper, an HM model is developed to evaluate the effectiveness of the proposed algorithms in simulations. The following scenarios are considered in the proposed HM model to closely emulate real traffic environments (such as highways and city streets).
1) A bidirectional four-lane street (two lanes in each direc-tion) is assumed in the HM model.
2) In the moving directions (i.e., MNs move along the x di-rection at the lower two lanes and along the−x direction at the upper two lanes), the MNs should maintain their speeds based on an average speed, which is prespecified for each lane.
3) The MNs may deviate within the lane for a certain degree in the lateral direction (i.e., along the y direction). 4) Lane changes are allowed in the HM model. If the speed
of the MN at the slow (fast) lane is larger (smaller) than a predefined value, the MN should start the lane-changing process.
The HM model that satisfies the above traffic scenarios can be formulated by modifying from the GMM model, as described in Section III-A. In the HM model, the motion equation of MNs along the x direction is represented by xk, as in (3). The
associated Vk is obtained from (2), where ¯V is selected based
on the average speed of each lane. The moving direction αk
should be confined to the lane direction with small variations due to different driving behaviors. αvcan be modified from (1)
by setting γ = 0 as
αk= ¯α + Xαk−1 (26)
where ¯α = 0 (or π) for moving along the positive (or negative) x direction. The zero-mean Gaussian-distributed random
vari-able Xαk−1is varied based on a preassumed driving deviation.
The lateral displacement ykcan be modified from (4) as
yk = y0+ Vk−1δt sin αk−1 (27)
where y0 is a constant value (i.e., y0= yslow for the slow
lane and y0= yfast for the fast lane), depending on the initial
position of the MN along the y direction. The reason of using a constant y0is to simulate the driving behavior of maintaining
the MN along the center of a lane.
Fig. 6 shows the state diagram for the lane-changing scenar-ios. The probability distribution function of the MN’s speed
Vk is denoted as fVk(v), where Vk is a Gaussian-distributed
random variable [as can be observed from (2)]. To satisfy the conditions for lane change, the following cumulative distribu-tion funcdistribu-tions (cdfs) are defined.
1) F1(v) = fVk[Vk > Vupper] =
∞
VupperfVk(v)dv: F1(v) is
the cdf when the MN’s speed Vkis greater than the upper
speed threshold Vupper. As shown in Fig. 6, the MN will
move to the fast lane if it is currently in the slow lane.
Fig. 6. State diagram for lane-changing scenario.
2) F2(v) = fVk[Vk< Vlower] =
Vlower
−∞ fVk(v)dv: F2(v) is
the cdf when the MN’s speed Vkis smaller than the lower
speed threshold Vlower. The MN will move to the slow
lane if it is currently in the fast lane. 3) F3(v) = fVk[Vlower≤ Vk ≤ Vupper=
Vupper
Vlower fVk(v) dv :
F3(v) is the cdf when the MN’s speed Vk lies between
the lower and upper speed limits. The MN will remain in its current state.
Both the RWM model and the HM model will be utilized in the simulations to evaluate the performance of the proposed algorithms in the following subsections.
B. Simulation Parameters
The simulations were conducted using the Network Sim-ulator (ns-2, [37]) to compare the proposed VAR, PMLAR, PMLAR-V, PMLAR-LV, and PMLAR-PV algorithms with the existing LAR-Box and DSR protocols. There are 70 MNs existing in the network with a total of ten constant bit rate connections in the simulations. The results are obtained by averaging the 100 simulation runs for 500 s. The average speeds of the MNs selected in the simulations are 5, 10, 15, and 20 m/s. The velocity thresholds that are utilized in (9) and (10) are assigned as δ1i= 0.125, δ2i= 0.5, δ1j = 0.25, and δ2j= 0.5. It is noted that δ1i and δ1j are selected smaller
compared with δ2i and δ2j since they will both be multiplied
by the velocity variables in the criterions. Moreover, δ1iand δ2i
are chosen such that ∆Viis greater than a positive value, which
indicates that the selected node Ni should move toward the
destination node D. The relative perpendicular speeds between
Ni and D (i.e., ∆Vj) should fall within the upper and lower
bounds as determined by δ1j and δ2j. It is also found in the
simulation results that the performance evaluation of the VAR-based schemes is not sensitive to these four velocity thresholds. The threshold µt is selected as π/2 for determining if the
DR or IR types of routing should be utilized. For the IR type of routing to happen (as shown in Fig. 5), µ > µt=
(πSD/2)/(SD) = π/2 indicates that the lower bound for the accumulated routing distances between the MNs is selected as half of the circle perimeter with diameter equal to SD, where
SD is the direct distance between S and D. To validate the
effectiveness of the proposed algorithms, the following three types of mobility models are utilized in the simulations with their parameters.
1) RWM model: The RWM model is adopted within the simulation area of 500× 500 m2. The MNs’ average pause time is set at 5 s.
2) RWM-IR model: The RWM-IR model is utilized in the simulations, as shown in Fig. 7. There exist an unreach-able area of 300× 300 m2within the simulation area of
Fig. 7. Simulation model for the IR scenario.
500× 500 m2. The IR type of routing as described in the
PMLAR protocol will be validated using this scenario. The MNs’ average pause time is also set at 5 s.
3) HM model: The HM model is adopted with four straight lanes (two lanes in each direction), each being 3.2 km in length and 4 m in width. The average speed of the MNs in the fast lanes is 4 m/s greater than that in the slow lanes. The variance of Xαk−1for each MN is chosen
to be 3◦.
C. Parameter Estimation
As mentioned in Section III-B, the parameters γi’s are
adapted online based on the RLS method [as in (17) and (18)]. To provide a fast converging RLS method, a proper selection of the parameter λ is required. In this subsection, the appropriate selection of λ is determined based on the different moving behaviors within the RWM and HM models.
1) Parameter Estimation for the RWM Model: As
men-tioned in Section V-A, the RWM model primarily consists of 1) a constant speed motion along the arbitrary direction and 2) a stationary speed motion for a prespecified time interval. The moving behaviors of the RWM model can be represented by the GMM model, which possesses both 1) high randomly selected moving angles [i.e., smaller values of γ1 in (1)] and
2) switching movements between constant speed and stationary motions [i.e., alternating between larger and smaller values of
γ2 in (2)]. Therefore, the behaviors of the RWM model can
be emulated by the GMM model that is utilized within the proposed schemes. The main purpose of this subsection is to offline obtain and validate the tunable parameter λ within the RLS method such that the RWM model can be emulated by the GMM model with its parameters ˆγiadapted online.
Several parameters in the GMM model are prespecified to offline emulate the corresponding RWM model. For validation purposes, the parameters in (1) are selected as (γ1, ¯α, Xαk−1) =
(a, 360· ζ · rand(0, 1), N (0, 10)) for t = {t ∈ |(ζ − 1) ·
ts2≤ t ≤ ζ · ts2, ζ > 0}. The parameter γ1 is assigned a =
(0.0, 0.1, 0.2) as the three different simulation cases shown in Fig. 8. The relative low values of γ1 are selected to reflect
that the moving angles αkmostly depend on the average angle
Fig. 8. Parameter estimation ˆγ1for the RWM model with λ = 0.9755 (solid
line: estimation for γ1= 0.2; dashed line: estimation for γ1= 0.1; and dotted
line: estimation for γ1= 0.0).
Fig. 9. Parameter estimation ˆγ2 for the RWM model with λ = 0.9755
[solid line: estimation for γ2= (0.1, 0.9); dashed line: estimation for
γ2= (0.2, 0.8); and dotted line: estimation for γ2= (0.3, 0.7)].
¯
α and the corresponding variations. The ζth ¯α is randomly
selected based on the different time interval ζ· ts2, where ts2
is chosen as 460 s in the simulations. As illustrated in Fig. 8, the three types of lines (i.e., solid, dashed, and dotted lines) represent different values of γ1to be estimated using the RLS
method. For example, the curved solid line stands for the estimated ˆγ1, which attempts to track the straight solid line
with γ1= 0.2. Several values of the parameter λ have been
experimented to offer better tracking ability for the validation models. It is found that λ = 0.9755 can provide satisfactory tracing performance compared with the other λ values.
As shown in Fig. 9, the parameters in (2) are selected as (γ2, ¯V , XVk−1) = ((b, c), (10, 0),N (0, 2)), where γ2= b, for (ζ− 1) · ts2≤ t ≤ ζ · ts1+ (ζ− 1) · ts2 c, for ζ· ts1+ (ζ− 1) · ts2≤ t ≤ ζ · ts2 . (28)
Fig. 10. Parameter estimation ˆγ1for the HM model with λ = 0.9999 (solid
line: estimation for γ1= 0.8; dashed line: estimation for γ1= 0.5; and dotted
line: estimation for γ1= 0.3).
It is noted that ts1= 400 s and ts2= 460 s are selected in the
simulations, as in Fig. 9. The three cases are represented by the variations of the parameter γ2= (b, c) = (0.9, 0.1), (0.8, 0.2),
and (0.7, 0.3). The value b of γ2 represents the time intervals
where comparably linear movement of the RWM model is assumed (with ¯V = 10 m/s), while γ2= c indicates the time
intervals that the MNs are relatively stationary ( ¯V = 0 m/s)
with less dependency to their previous moving behaviors. The selection of the two different values of γ2 is attempting to
emulate the movement of the RWM model, i.e., the MNs move in constant speeds and then pause for a certain time period. It is also evaluated in the simulations that λ = 0.9755 (as shown in Fig. 9) provides acceptable tracking performance compared with the other values. The value of λ = 0.9755 will be utilized as the simulation parameter, while the RWM model is used as the underlying mobility model in the simulations.
2) Parameter Estimation for the HM Model: Similar to the
discussion in the previous subsection, the parameter λ within the RLS method is acquired offline using the GMM model to emulate the corresponding HM model. The parameters in (1) are selected as (γ1, ¯α, Xαk−1) = (a, 0,N (0, 10)). The
parame-ter γ1 is chosen as a = (0.3, 0.5, 0.8) as the three different
simulation cases, as in Fig. 10. The values of γ1 are selected
to reflect the moving angles that are selected depending the driving behaviors, which can either be randomly selected (i.e.,
γ1= 0.3) or linearly related to previous driving manners (i.e., γ1= 0.8). It is found in the simulations that λ = 0.9999 is
sufficient for the estimated ˆγ1 to trace the various γ1 values,
i.e., for γ1= 0.8, 0.5, and 0.3, as illustrated in Fig. 10.
The parameters in (2) are selected as (γ2, ¯V , XVk−1) =
(b, ζ· rand(5, 25), N (0, 2)) for t = {t ∈ |(ζ − 1) · ts2≤
t≤ ζ · ts2, ζ > 0}. The parameter γ2 is assigned b = (0.7,
0.8, 0.9) as the three different cases, as illustrated in Fig. 11. The relatively high values of γ2 are selected to reflect that
constant speeds are preserved under highway driving. The average speed of each MN is changed at certain time instants, where the ζth ¯V is randomly chosen based on the different
time interval ζ· ts2. The value of ts2is also selected as 460 s
Fig. 11. Parameter estimation ˆγ2for the HM model with λ = 0.9999 (solid
line: estimation for γ2= 0.9; dashed line: estimation for γ2= 0.8; and dotted
line: estimation for γ2= 0.7).
in the simulations. As shown in Fig. 11, adequate tracking performance can be obtained for the estimated ˆγ2by selecting λ = 0.9999, which is the same value as in the selection of ˆγ1
for the HM model. The value of λ = 0.9999 is adopted as the parameter, while the HM model is utilized as the underlying mobility model in the simulations.
D. Simulation Results
The following four metrics are considered for performance comparison:
1) The Data Packet Delivery Ratio: the percentage of suc-cessful deliveries for the data packets;
2) The End-to-End Delay: the average time elapsed for de-livering a data packet from the transmitter to the receiver; 3) The Control Packet Overhead: the ratio from the total transmitted control packets to the total received data packets;
4) The Route Life Time: the average life time of a route for data transmission.
With the simulation parameters and the performance compari-son metrics stated above, the simulation results are explained in the following two subsections.
1) Performance Comparison Between the VAR-GMM and the VAR-CSM Algorithms: In this subsection, the performance
comparison for the VAR algorithm using either the GMM or the CSM model is evaluated. Figs. 12–14 show the performance comparison for both VAR-GMM and VAR-CSM algorithms that are implemented under the RWM and HM models (with
V = 5, 10, 15, 20 m/s). It can be observed that the VAR-GMM
algorithm outperforms the VAR-CSM scheme under both RWM and HM models in all the three comparison metrics.
As illustrated in Fig. 12, the data packet delivery ratio of the VAR-GMM algorithm is around 2%–3% higher than that of the VAR-CSM scheme under both mobility models. The end-to-end delay of the VAR-GMM algorithm is around 40–50 ms less, compared with the VAR-CSM scheme at V = 20 m/s (as in Fig. 13). The control packet overhead of the VAR-GMM
Fig. 12. Data packet delivery ratio versus average velocity for both VAR-GMM and VAR-CSM algorithms under RWM and HM models.
Fig. 13. End-to-end delay versus average velocity for both VAR-GMM and VAR-CSM algorithms under RWM and HM models.
scheme is around 1.5–2 smaller than that using the VAR-CSM algorithm for both mobility models at V = 20 m/s (as in Fig. 14). It is noticed that the simulation results as above can be expected since the more accurate is the predicting model used (i.e., the GMM model), the better can the routing performance be obtained. However, the benefit of using the CSM model is that it provides efficiency in algorithm com-putation compared with the GMM model. It is also observed that the simulation results obtained by using the HM model illustrate better performance compared with that from the RWM model in all three metrics. These results are obtained because the selected simulation scenario for the HM model has better routing efficiency compared with that from the RWM model. In the next subsection, the GMM model is selected as the predicting mechanism for the performance evaluation of the proposed algorithms, including the VAR, PMLAR, PMLAR-V, PMLAR-LV, and PMLAR-PV algorithms.
Fig. 14. Control packet overhead versus average velocity for both VAR-GMM and VAR-CSM algorithms under RWM and HM models.
2) Performance Evaluation of the PMLAR and Its Associ-ated Algorithms: In this subsection, the performances between
the proposed VAR, PMLAR, PMLAR-V, PMLAR-LV, and PMLAR-PV algorithms, and the existing DSR and LAR-Box protocols are compared. It is noted that the GMM model is utilized as the prediction model in all the simulation cases. Figs. 15–18 illustrate the performance comparison under the RWM and RWM-IR models, while the simulation results under the HM model are shown in Figs. 19–22. As seen in Fig. 15, the data packet delivery ratio for the PMLAR-PV algorithm can achieve around 8% and 13% higher ratios than that from the LAR-Box algorithm under the RWM and RWM-IR models (at speed = 20 m/s). The PMLAR-derived protocols (i.e., the PMLAR, PMLAR-V, PMLAR-LV, and PMLAR-PV schemes) can acquire a higher packet delivery ratio than the VAR, LAR-Box, and DSR algorithms. It can also be found that the separation between the PMLAR derivative schemes and the non-PMLAR protocols becomes more obvious under the RWM-IR model (as shown in the right plot of Fig. 15). The reason comes from the additional waypoint information that is obtained from the location service while the PMLAR-related schemes are encountering the IR scenarios. This waypoint information helps to direct the route discovery and packet-forwarding processes in the right directions.
Fig. 16 illustrates that the end-to-end delay of the PMLAR-PV scheme is around 180 and 210 ms less, compared with that of the LAR-Box protocol under the RWM and RWM-IR models (at V = 20 m/s). It is also worthwhile to notice that the PMLAR scheme can achieve a smaller end-to-end delay compared with the PMLAR-V, PMLAR-LV, and PMLAR-PV algorithms under the RWM-IR model (as shown in the right plot of Fig. 16). It is found in the simulation data that the selection of forwarding MNs based on the relative velocity (i.e., in the PMLAR-V, the PMLAR-LV, and the PMLAR-PV schemes) causes an adverse effect under the RWM-IR model. The selected forwarding MNs that move in the same direction as the destination node may move out of the regions confined
Fig. 15. Packet delivery ratio versus average velocity with the RWM (left plot) and RWM-IR (right plot) models.
Fig. 16. End-to-end delay versus average velocity with the RWM (left plot) and RWM-IR (right plot) models.
by the waypoints, which can result in potential link breakage for the delivering routes. Nevertheless, the PMLAR-derived schemes still outperform the DSR, LAR-Box, and VAR pro-tocols with smaller end-to-end delay.
Due to the mechanism of proactive route maintenance, the PMLAR-PV scheme contributes more control overhead than the PMLAR, PMLAR-V, and PMLAR-LV algorithms, as shown in Fig. 17. The destination node D periodically monitors its own position to ensure that it will stay within the Predicted
Zone. If D moves close to the boundary of the region, it
will send an update information back to the source node S, and a new route discovery process will be initiated by S. The information update packets and the new route discovery packets are the causes of additional control overhead. However, the total control overhead of the PMLAR-PV scheme is still lower than that obtained from the DSR, LAR-Box, and VAR protocols. Fig. 18 shows the performance comparison under the average route lifetime metrics. The proposed PMLAR-PV
scheme can provide additional 24 and 28 s of average route lifetime compared with that of the LAR-Box algorithm under the RWM and RWM-IR models (at V = 20 m/s). It is also observed that comparably robust communication linkages can be obtained using the PMLAR derivative schemes, especially under the RWM-IR model.
From Figs. 15–17, it can be observed that the performance of the PMLAR-LV scheme, which adopts a local route repair, is comparably better than the PMLAR-V algorithm, which utilizes the end-to-end route repair. The localized mechanism decreases the delay time while repairing a route and provides a higher probability for a successful packet delivery. How-ever, it is worthwhile to find that the routes created by the PMLAR-LV scheme tend to become unstable in the high-speed environments. As shown in Fig. 18, the route lifetime of the PMLAR-LV scheme is higher than that from the PMLAR-V method under lower speed environments. However, the decreas-ing rate from the PMLAR-LV method is comparably larger than
Fig. 17. Control packet overhead versus average velocity with the RWM (left plot) and RWM-IR (right plot) models.
Fig. 18. Route life time versus average velocity with the RWM (left plot) and RWM-IR (right plot) models.
that from the PMLAR-V scheme as the velocity is increased. The primary reason can be attributed to the inferior local route repair under dynamic moving environments. As the speeds of the MNs are increased, the communication links between the MNs are inclined to be broken. The PMLAR-V scheme offers the end-to-end route repair, which has a higher possibility to acquire a more stable route for packet delivery. On the other hand, the PMLAR-LV method provides a localized view for route repair, which may tend to become unstable under high-speed environments. In general, the route lifetime obtained by the PMLAR-PV scheme still outperforms all the other algorithms under various velocities due to its proactive route maintenance.
Moreover, since the LAR-Box protocol can be regarded as the special case of the PMLAR scheme (i.e., by assigning
γi= 1 in the GMM prediction model), the PMLAR algorithm
and its associated schemes provide better routing performance compared with the LAR-Box protocol. It can also be observed
that the performance of the LAR-Box protocol is inferior under the RWM-IR model compared with that under the RWM model (as in Figs. 15–18). Due to the unreachable area within the RWM-IR model, the source node will not be able to construct routes within the rectangular box (i.e., the Request Zone) gener-ated by the LAR-Box protocol. The unreachability problem can be resolved by adopting the additional waypoint information within the PMLAR-related schemes, which can acquire better performance under the three metrics.
In the HM scenario (as in Figs. 19–22), the velocity-based protocols (i.e., the VAR, PMLAR-V, PMLAR-LV, and PMLAR-PV schemes) outperform the nonvelocity-based al-gorithms (i.e., the PMLAR, LAR-Box, and DSR schemes) in the packet delivery ratio, end-to-end delay, and route lifetime metrics. The similar moving behaviors of the MNs within the HM model cause the velocity-based algorithms to provide better routing efficiency compared with the other schemes. Moreover, the PMLAR-LV method provides a better route
Fig. 19. Packet delivery ratio versus average velocity with the HM model.
Fig. 20. End-to-end delay versus average velocity with the HM model.
lifetime compared with the PMLAR-V scheme under differ-ent speeds. Due to the similar moving behaviors within the HM model, the localized route repair is sufficient to acquire feasible routes in contrast to the end-to-end manner. It can be observed from the simulation results that the PMLAR-PV scheme can still preserve adequate routing performance, i.e., with around 11% higher packet delivery ratio and 90 ms less end-to-end delay compared with the LAR-Box protocol (under the speed of 20 m/s).
Similar results can be obtained in the control overhead metric, as in Fig. 21. The PMLAR-PV algorithm contributes a slightly higher control packet overhead (compared with the PMLAR-V, PMLAR-LV, and PMLAR schemes) due to its proactive route maintenance. It can also be seen that the control packet overhead acquired from the VAR protocol is higher than that from the LAR-Box protocol. The higher overhead comes from the whole-region flooding by the VAR algorithm, while
Fig. 21. Control packet overhead versus average velocity with the HM model.
Fig. 22. Route life time versus average velocity with the HM model.
the LAR-Box protocol incurs a comparably smaller overhead by constraining its flooding area within the Request Zone.
VI. CONCLUSION
In this paper, the VAR and PMLAR protocols have been proposed for MANETs. The VAR algorithm considers the relative velocity between the intermediate node and the desti-nation node in its protocol design. The PMLAR algorithm is designed with the incorporation of the predictive moving be-haviors of the MNs. The prediction mechanism proposed in the PMLAR algorithm can effectively forecast the future trajectory of the destination node. The PMLAR-V scheme enhances the PMLAR protocol with the VAR algorithm, while the local route repair is adopted within the PMLAR-LV scheme. The PMLAR-PV algorithm further encompasses a proactive main-tenance scheme in the protocol design. By incorporating the
![Fig. 6 shows the state diagram for the lane-changing scenar- scenar-ios. The probability distribution function of the MN’s speed V k is denoted as f V k (v), where V k is a Gaussian-distributed random variable [as can be observed from (2)]](https://thumb-ap.123doks.com/thumbv2/9libinfo/7987467.159340/9.891.461.831.102.191/changing-probability-distribution-function-gaussian-distributed-variable-observed.webp)
