Network Models For Seat Allocation On Flights: Moshe Dror,?
Network Models For Seat Allocation On Flights: Moshe Dror,?
Network Models For Seat Allocation On Flights: Moshe Dror,?
00 + 03
Printed in Great Britain. 0 1988 Pergamon Press plc
and
SHAUL P. LADANY
Industrial Engineering and Management, Ben Gurion University of The Negev, Israel
Abstract-This paper examines the problem of proper (optimal) control over the seat allocation
on flights. Given a heterogeneous fleet of aircraft types, multi-leg flights, a number of different
passenger categories, and cancelations, an airline’s objective is to devise an effective system
which aids in setting the seat allocation targets for each category of passengers on each flight.
This issue is analyzed by a number of authors in the context of economic, simulation based,
probabilistic, and mathematical programming studies. We present an attempt to address this
problem from the systems prospective emphasizing characteristics such as: passenger cancelations,
multi-leg flights, and rolling tactical planning time horizon. Starting from a very simple network
flow model for a single flight with a number of intermediate stops, a number of progressively
complex models are presented. The airline flights and the seat allocation system are represented
as a generalized network flow model (with gains/losses on arcs) with the objective of flow
maximization (profit maximization). This modeling approach does not claim to replace the seat
allocation approaches presented in Alstrup et al. (1985), Mayer (1976). Richter (1982), Simpson
(1985a). and Wang (1983), but rather construct seat allocations utilizing some of those referenced
schemes in a parameter setting mode for a large network model. The objective of this paper is
not to report on computational experiments, but to present a modeling approach which seems
to be promising, if somewhat speculative.
INTRODUCTION
Over the years, airline companies have practiced various promotional pricing policies
and devised innovative low fares to develop and stimulate consumer demand and to
fulfill the needs of new market segments. As Wang (1983) points out, this price differ-
entiation was designed to set up barriers to prevent the high-yield (business) passengers
to switch to the lower-priced promotional fares. Given the passenger price differentiation,
the airlines in turn had to decide on an appropriate seat allocation on each flight for the
subsequent passenger categories. This was especially important since the low-yielding
fares tended to block seats for the higher-yielding business passengers. Additional block-
ing phenomena occur in the case of the multi-leg flight reservations where the long-
haul passengers are often unable to obtain seats because they are blocked by the shorter-
haul passengers. For large commercial airlines, efficiently setting and updating seat
allocation targets for each passenger category on each flight is an extremely difficult
problem. In addition, the airlines experience cancelations and “no-shows” (unused
reservations) for almost all of their flights.
In most of the seat allocation studies originated in the airline industry, the issue of
cancelations and no-shows is not addressed (Mayer, 1976; Richter, 1982; Simpson, 1985a
and 1985b; Titze and Griesshaber, 1983; Wang, 1983). Moreover, those studies and
others (Hersh and Ladany, 1978; Ladany and Bedi, 1977) tend to examine the seat
allocation problem from the local prospective of one flight at a time type analysis. Mayer
(1976) examines the simple and well-known models for seat allocation on a single flight
and performs sensitivity analysis of these simple formulas. In Hersh and Ladany (1978)
and Ladany and Bedi (1977), the seat allocation problem for a single flight is analysed
from the probabilistic perspective and optimal seat allocation is derived incorporating
allocation terminology will be clarified as the paper progresses.) The final network model
is to be resolved (partially or fully) at the end of each time segment producing new
optimal seat allocation goals. The lower and upper bounds on flows together with the
gain/loss coefficients are to be updated automatically at each time segment based on
the historical data and the present reservations status.
Basically, given a potential passenger who has indicated the destination, fare cat-
egory, and preferred time of departure, the airline (directly or through a travel agent)
either accepts a reservation or offers the passenger flight alternatives. Any subsequent
reservation for this passenger has a different likelihood of cancelation/no-show. All
those factors should be incorporated in the seat allocation model, together with the
average expected contribution to profit for each of the alternatives selected by the
passenger. These alternatives could differ in departure time, aircraft type, and whether
the flight is direct, with intermediate stops or even with switchovers.
Next we present the appropriate network flow models.
The single flight with intermediate stops seat allocation problem is modeled as a
maximal flow problem on the following network (Fig. 1).
The node s is the source-node and the node c is the sink node. They correspond to
flight’s origin and final destination respectively. The nodes, cl and c,(i = 1, 2, . . . , k),
represent category i of passengers. More precisely, the arc (cr, ci) represents category
i of passengers. The nodes d,(i = 1,2, . . . , I) represent the different intermediate stops
for the flight. The arc (t’, t) represents the aircraft used for that flight. The upper bound
(capacity) on the arc (t’, t) represents the maximal passenger capacity for this aircraft.
The nodes Ci, (i = 1, 2, . . . , k) are connected with the nodes dj, (j = 1, 2, . . . , I)
according to the destination of the different passenger categories.
The arcs (s, cl) have no flow restrictions (lower-upper bound), no cost/profit weight
on the flow, and no gain/loss flow coefficient. Clearly, originally there is no restriction
on the number of passengers who might belong to a certain category (infinite population
assumption). The arcs (c,!, c;) have only lower and upper bounds on their flow reflecting
for that category of passengers their present and maximal (expected or restricted) res-
ervation levels respectively. Arcs (ci, dj) are the only arcs with gain/loss flow coefficients
(which correspond to 1 minus the expected fraction of cancefations and no-shows) and
a positive (might also be negative) weight representing the average expected profit for
the corresponding passenger category. Those arcs do not have upper bounds on their
flow, only lower (zero) bound which is common to all arcs in this and subsequent models.
The arcs (& t!) have upper and lower bounds on their flow. Those bounds represent
allocation limits relative to a specific flight destination. The arc (t’, I) has an upper
bound on the flow identical with the aircraft’s capacity. By solving the maximal weight
flow problem on this directed network we obtain the seat allocation on this flight which
maximizes the average expected profit.
The solution to this problem on the given network is trivial; Maximize the flow
through the path from node s to node t with the maximal profit weight. Once this path
is saturated, go to the next profitable path and again maximize the weighted flow, etc.
The arc weight (profit) exists only on arcs (ci, dj) and the paths from node s to node t
through these arcs are uniquely determined.
The profit weight on an arc type (Ci, dj) represents the average expected profit
contribution of a passenger from category ci flying to destination dj. The gain/loss
coefficient on such arc represents 1 minus the expected fraction of cancelations and no-
shows for the particular category of passengers. This expected fraction of cancelations
and no-shows will vary with time and reservation status before the flight as noted already
in the literature.
Illustration
Assume an aircraft (Boeing 767) of 179 passenger capacity and three basic passenger
categories: first class, second class, and third class. The seats on this aircraft are (usually)
partitioned into three sections: 10 seats for the first class, 50 seats for the second class,
and 119 seats for the third class. Assume a flight with one intermediate stop and no
passenger boarding at that stop. In this case we have 6 passenger categories denoted
c11, Cl27 c217 czz, c31, and ~32 indicating the class and the destination. Assume that the
average expected profit from passenger flying first class to the final destination is 4 units
(a simplified measure), for passenger in c iI category it is 3 units, for passenger in cZ2
category it is 2 units, 1.5 units for passenger in cZ1, 1.5 units for passenger in c3*, and 1
unit of profit from passenger flying to the intermediate stop. In addition, assume the
same fraction of cancelations and no-shows for all the passenger classes at 0.10. The
gain/loss coefficients (1 minus the cancelations and no-shows fraction) on arcs (ci, dj)
and the weights on those arcs are now known. The upper bound on the flow through
the arc (t’, t) is 179 (the capacity of the aircraft). The lower bounds on arcs (s, cl) are
given by the present reservation levels. The upper bounds are either fixed by specific
managerial policy (such as 10 seats, 50 seats, and 119 seats for classes 1, 2, and 3
respectively) or calculated accounting for the expected reservation levels given present
status. The basic seat partition might fix some upper (and lower) bound parameters if
upgrading (i.e., assigning passengers from the second-class to first-class seats) (or down-
grading) of passengers is not permitted. The solution in this case is obtained in straight-
forward fashion by allocating seats according to their expected contribution to profit.
The problem gets more complicated when the passenger demand for the flight exceeds
the flight capacity and there exists a need to offer passengers alternative flight selections.
The cancelation fractions need not be uniform (for business passengers the cancelation
and no-show rate might reach 0.40 on some flights) and the average expected profit for
passenger category might be modified accounting for boarding passengers in the inter-
mediate stop. When incorporating more realistic problem parameters, the corresponding
network is not any more simple and neither is the solution scheme. These complications
are incorporated in the next model.
Note: The above basic network model does not accomplish, so far, anything new.
The arc parameters such as bounds on flow, average expected profit margin for passenger
category, and gain/loss coefficients are computed separately. The optimal seat allocation
for a single multi-leg flight can be easily obtained through other known techniques. The
contribution and novelty of this basic network model emerges in the larger interconnected
network models which incorporates many such basic network units and many time pe-
riods. This will hopefully become more transparent in the next section.
Seat allocation on flights 243
THE FIRST NETWORK MODEL
In this first network model presented below, we interconnect many flights (basic
network models) into one single network. This network’s time span is the entire (short-
term) planning period (say four or six weeks). The seat allocation targets are obtained
by solving the weighted maximal flow problem on the network. This solution procedure
is repeated at the end of each time segment (i.e., daily) with updated network parameters.
Passengers in all categories will be offered (directed to) a choice in reservation selection
according to the solution obtained in the last run. The figurative network representation
of the first network model is illustrated in Fig. 2 for one flight over five time segments.
The notation (specifically the indices) is slightly different than the one for the basic
model in order to incorporate the additional complexities.
The node d, in Fig. 2 is an artificial node representing all the future flights beyond
the planning period. Given that the sum of the upper bounds on the flow for some of
the flights might exceed the flight capacity of the aircraft (i.e., the demand for a flight
exceeds its capacity), the optimal allocation of seats (and diversion of passengers) for
the different passenger categories is no longer a trivial problem solved by inspection.
The model of course is of no use if all the flights have a low load (demand). Clearly,
Fig. 2. Five-flight network: Three customer categories with two identical destinations for each flight.
244 M. DROR et al.
the model can be also used for re-figuration studies to estimate the revenue impact of
additional seats on various routes.
Next, we present the corresponding mathematical formulation for the first network
model and explain the notation.
where
is the average expected profit contribution to flight 1of one passenger from category j who originally
selected flight i and whose destination is d,.
is the fraction of passengers with destination d, expected not to cancel if in category c,, and scheduled
for flight 1. One may expect that r:r 2 rf for any 1 f i.
is the flow (the number of reservations) that should be allowed on the arc from c,, to d,,.
Constraints. f(x, y) 5 f(x, y) zs u(x, y) for all arcs (x, y) in the network as indicated.
u(x, y)-is the upper bound on the flow from node x to node y and Z(x, y) is the lower
bound on the flow from node x to node y.
At this point we note that this network optimization model is not restricted to
identical planes or identical flights from an origin to a final destination. The precise
indication as to the type of aircraft on a given flight is via the upper bound on the flight’s
capacity, i.e. the upper bound on the flow through the arc (t;, ti). Different flights with
even one overlapping intermediate stop (which could be a final destination for one flight)
can also be incorporated in this flow model. Needless to say, this model can also be
implemented for a bus company with long routes and intermediate stops.
In order to clarify the above network flow model, we examine the interrelation
between the lower and upper bounds on the different arcs in the model.
The lower bound on the flow through the arc type (ck, cij) indicates the present
reservation level for the passenger-category j whose initial choice was the flight i. The
upper bound on the flow through that arc indicates either the expected demand for that
passenger category and flight given the present reservation status, or some strategic/
managerial restriction on reservation for a certain (low-yield) category of passengers.
The lower bound on the flow through the arc type (C,, dip) represents the present
reservation status on flight I for passengers of the jth category whose initial choice was
flight i and whose destination is dp. These lower bounds multiplied by the noncancelation
Seat allocation on flights 245
fraction and summed for a specific 1, p over all i, j’s in the connecting arcs are the lower
bounds on the arc (d,, r;). The upper bound on the flow through arc (d,, r;) can
represent the expected demand for the destination d, on flight 1. The lower bound on
flow through the arc (f;, f,) is the sum of the lower bounds on the arcs (d,p, ti) and the
upper bound is the actual number of seats on the flight p.
This network flow model does not explicitly permit the possibility of passengers
boarding a flight at an intermediate stop. This eventuality can be incorporated implicitly
into this model by properly adjusting the average expected profit margin for the passenger
flying only from the origin to an intermediate stop by the average expected additional
profit for that plane seat in a flight-leg from that intermediate stop to the final destination.
One major drawback of this first network model is its inability to incorporate the
case in which passengers are allowed to switch planes at an intermediate stop in order
to arrive at their final destination. It is an important aspect to be considered when setting
allocation goals for seats on flights, since approximately 50% of the passengers switch
flights at an intermediate stops.
In order to include the possibility for passengers to switch planes at an intermediate
stops, a more complex version of the network flow model is required. This quite complex
network flow model is the topic of the next section.
departing
passenger
intermediate stop
passengers
+
departlnq
pa’ssengek
Fig. 3. A schematic presentation of two flights with one switch-over intermediate stop.
green
destination
black
origin
destination
s : source node
I : sink node
(c;. Ci) : jth category of passengers on flight i
(Q ,Cll : jth category of passengers on flight i but boarding at an intermediate stop.
d, : denotes a node which corresponds to the destination d, which is final for some passengers of flight
i departing at an intermediate stop,Note that destination d, may be common to many flights, for
instance flights a and b will have d,, d,,.
d$ : denotes a node which corresponds to destination d, in flight i, and is related to passengers entering
the flight after the kth flight leg (k t 0).
aa: denotes a node which corresponds to a switch-over in flight i, and is related to passengers entering
the flight i after the kth flight leg and traveling to destination d,.
lower-bound constraint on the flow from node x to node y
upper-bound constraint on the flow from node x to node y
fraction of passengers with destination d, expected not to cancel if in category c,, and assigned
flight 1, boarding after the kth flight leg and making no switch-over thereafter.
fraction of passengers with destination d, expected not to cancel if in category c,, and assigned
flight I, boarding after the kth flight leg and switching over later to eventually reach destination
d
(I;, t,-) : d:notes the jth flight-leg of flight i.
i,i : denotes the final destination of flight i occurring on its jth stop.
;j = {r, k such that the arc (c,, d$) exists, where d, is destination of c,,}.
I, = {I, k such that the arc (c,~, d$) exists, where dp is the destination of ci,}.
$4.= {i, j such that the arc (q. df,,) exists, where d, is the destination of cii}.
J, = {i, j such that the arc (cii, d&) exists, where d, is the destination of c,,}.
K, = {m. n such that the arc (t,_, d$) exists}
R, = {m, n such that the arc (Imn, (if) exists}
The notation (above) is aimed at representing in a detailed manner all the intricaties
of this complex version of the seat allocation problem. We did not aim to present the
most concise (in terms of variables) representation of the problem, but rather attempted
to be descriptive.
The objective function.
Constraints.
For the flow constraints below, the symbol in the left margin indicates the node for
which the conservation of flow is about.
The next two equations express equal flow requirements between (specific) non-
adjacent arcs. The equations are referred to as “switch-over” and “out”, respectively.
switch-over:
out:
f($,, &) = 2 f(df;‘, t;J for all 1, p where d, is the pth destination on flight 1.
1515p
We do not claim that the above mathematical model can be easily solved. The model
presents a very detailed (though still partial) and complex representation of the routine
activities for a commercial airline. In order to solve this type of a model for a large
operation, in our estimation, new solution techniques will have to be examined and
developed. Most of the necessary methodology exists and would have to be suitably
pieced together into an appropriate solution procedure.
Figure 5 presents the network flow illustration for the mathematical description of
this second (expanded) network flow model. Three flights are represented (1, 2, and 3),
originating at three different cities (black, brown, and white), and a total of six desti-
nations (yellow, blue, orange, green, red, and white). To simplify, each category of
passengers of a given flight is going to a different destination. Category Clj of passengers
illustrates the possibility of multi-switch-overs (two in this case) starting in flight 1,
transferring to flight 2 in the yellow city, and finally leaving flight 2 in the white city for
flight 3 where it will reach the desired destination, red city.
The objective for the airline is to construct an operating policy for the allocation
of seats to passengers flying full flights and flight-legs such that the expected total
contribution to profit for the entire flight schedule over a planning period is maximum.
Since the demands and cancelations for the full span and the intermediate span are
stochastic variables dependent on the time prior the flight departure and other factors,
it becomes necessary to derive the allocation policy as a sequential decision process. At
the start of each tth decision period, prior to the departure, a decision must be made
about the maximum additional number of reservations to be allowed for the different
passenger categories in the different flights and flight-legs. This decision must be based
upon:
(i) The aircrafts’ capacities. In this problem the aircraft’ capacities for the different
flights can be varied in order to examine the relationship between the maximum flow
(i.e., maximum profit solution) and the aircraft assignment to the different flights.
(ii) The average expected contribution to profit of a passenger from a given category
on a given flight-leg.
(iii) The probab’l’11st ic pattern of future demands, cancelations, and no-shows, for
each category of passengers and flight-legs.
(iv) The distribution of standbys/waiting list passengers for each flight leg.
(v) The present status of reservation for each passenger category and flight-leg.
(vi) The penalty cost incurred for passengers with confirmed reservation but no
seat available (overbooking cost).
CONCLUSION
most of the other approaches to this problem, we presented a systems approach which
examines the seat allocation for an entire fleet of aircraft types over a planning period.
We hope to have demonstrated that this modeling approach is quite flexible in terms of
the planning period and the flights as well as a large number of passenger categories.
The model could be implemented in relatively simple cases (a subset of flights),
where no switch-over passengers are allowed, in a rather straightforward mode of solving
the maximal weight flow problem on a network with gains/losses. When incorporating
the switch-over passengers into the model, the corresponding network evolves into a
more complex multicommodity network flow problem to which additional constraints
are added.
There are two major aspects of the present models which have not been discussed
in this paper: (i) the calculation of the parameters for this type of networks, and (ii) the
specifics of the computerized procedures (the algorithms) which can solve these network
models. Those issues require a separate forum.
REFERENCES
Alstrup J., Boas S., Madsen 0. B. G. and Vidal R. V. V. (1985) Booking policy for flights with passenger
types, IMSOR. The Technical University of Denmark, Research Report No. 01.
Hersh M. and Ladany S. P. (1978) Optimal seat allocation for flights with intermediate stops. Camp. Oper.
Res. 5, 31-37.
Kennington J. (1978) A survey of linear cost multicommodity network flows. Oper. Res. 26, 209-236.
Ladany S. P. and Bedi D. N. (1977) Dynamic booking rules for flights with intermediate stops. Omega S(6),
721-730.
Mayer M. (1976) Seat allocation, or a simple model of seat allocation via sophisticated ones. Proceedings of
the AGIFORS Sixteen Annual Symposium, Miami, Florida, 103-125, September.
Phillips D. T. and Garcia-Diaz A. (1981) Fundamentals of Network Analysis. Prentice-Hall, Englewood Cliffs,
New Jersey.
Richter H. (1982) The differential revenue method to determine optimal seat allotments by fare types. Pro-
ceedings of the AGIFORS Twenty-Second Annual Symposium, Longonissi, Greece, 339-353, October.
Simpson R. W. (1985a) Setting optimal booking levels for fhght segmentiwith multi-class, multi-market traffic.
Proceedinas of the AGIFORS Twentv-Fifrh Annual Svmoosium. Hikone. Jaoan. 263-271. October.
Simpson R. W. (1985b) Theoretical concept for capacity/yield management. Proceedings of rhe AGIFORS
Twenty-Fifth Annual Symposium, Hikone, Japan, 281-293, October.
Soumis F. (1979) Planification dune flotte d’avions. Ph.D. Thesis, Departement d’informatique et de recherche
operationnelle, Universite de Montreal, Publication #133, Centre de recherche sur les transports, Uni-
versite de Montreal.
Soumis F., Ferland J. A. and Rousseau J. M. (1979) MAPUM-a model for assigning passengers to flight
schedule. Publication #142. Centre de recherche sur les transports, Universite de Montreal.
Titze B. and Griesshaber R. (1983) Realistic passenger booking behaviour and the simple low-fare/high-fare
seat allotment model. Proceedinas of rhe AGZFORS Twenfv-Third Annual Svmposium., . Olive Branch.
Mississippi, 197-223, October. - ’
Wang K. (1983) Optimum seat allocation for multi-leg flights with multiple fare types. Proceedings of the
AGIFORS Twenty-Third Annual Symposium, Olive Branch, Mississippi, 225-246, October.