Displacement_problem_and_dynamically_scheduling_ai
Displacement_problem_and_dynamically_scheduling_ai
net/publication/245281259
CITATIONS READS
167 746
4 authors, including:
SEE PROFILE
All content following this page was uploaded by Mohan Krishnamoorthy on 10 July 2014.
www.palgrave-journals.com/jors
(3) D(X, x) is a separable function into a summation landing of an aircraft, and the landing of all successive
of terms involving only Xi and xi, so that the contri- aircraft, are respected. In order to formulate the problem
bution to total displacement of a change in decision define:
variable i, from Xi to xi, can be identified separately
P the number of aircraft
from the effect of all other changes in decision
Ei the earliest landing time for aircraft i
variables.
Li the latest landing time for aircraft i
Let: Ti the target (preferred) landing time for aircraft i
gi the penalty cost (X0) per unit of time for landing
Di(X, x) represent the contribution to total displace-
before target Ti for aircraft i
ment of a change in decision variable i from
P hi the penalty cost (X0) per unit of time for landing
Xi to xi (eg if D(X, x) ¼ N 2
i¼1 (Xi–xi) then
2 after target Ti for aircraft i
Di(X, x) ¼ (Xixi) )
Sij the required separation time (X0) between aircraft i
pi be the objective function weighting (X0)
landing and aircraft j landing (where aircraft i lands
per unit of displacement in variable i
before aircraft j)
(i ¼ 1,y,N)
xi the landing time (X0) for aircraft i, a decision
Di be the maximum displacement (X0) that
variable
we are prepared to accept in variable i
dij ¼ 1 if aircraft i lands before aircraft j
(i ¼ 1,y,N)
¼ 0 otherwise
Then the general form of the displacement problem is:
then the (single runway) static ALP has
minimise X
P
X
N ð1Þ ZðxÞ : ðgi max½0; Ti xi þ hi max½0; xi Ti Þ ð5Þ
lcost Z ðxÞ þ ldisp pi Di ðX; xÞ þ lmax Dmax i¼1
i¼1
Here, Z* and C* represent the original objective and Ei pxi pLi ; i ¼ 1; . . . ; P ð8Þ
constraints (Z and C) but amended to reflect any changes Equation (6) ensures that for each pair of aircraft one lands
that have occurred (eg the addition of new decision before the other, Equation (7) enforces separation between
variables). In the objective function, Equation (1), lcost, aircraft and Equation (8) ensures that each aircraft lands
ldisp and lmax (X0) are the weights (respectively) attached to: within its time window. The first/second maximisation terms
total cost of solution Z*(x); total cost of displacement in Equation (5) account for aircraft that land before/after
PN
i¼1 piDi(X, x); and maximum displacement Dmax. Equation target. This cost function is illustrated diagrammatically in
(3) limits the displacement for any variable, while Equation Figure 1. Colloquially gi and hi are the slope of the cost
(4) defines the maximum displacement Dmax. function before and after the target time Ti respectively.
Extending the above single runway formulation to the
multiple runway case is easily done and, for reasons of space,
Aircraft landing
will not be presented here. Note here that given a fixed
In this section, we first briefly review the aircraft landing sequence (equivalently values for the zero–one
landing problem (henceforth ALP) and then go on to variables dij above), an effective approach to deciding
define an example displacement function for the dynamic optimal landing times that minimise cost with respect to
ALP. that given sequence is to solve the LP that results from
Equations (5)–(8) when the zero–one variables are elimi-
nated. For convenience, we refer to this LP as ALPF
Aircraft landing problem
indicating that it is the ALP with a fixed landing sequence.
The ALP is the problem of deciding a landing time on Beasley et al1 first formulated the static ALP as a mixed-
an appropriate runway for each aircraft in a given set integer zero–one linear program and solved it numerically
of aircraft such that each aircraft lands within a predeter- for a number of test problems involving up to 50 aircraft and
mined time window; and separation criteria between the four runways. Beasley et al2 studied the single runway static
56 Journal of the Operational Research Society Vol. 55, No. 1
ALP and presented a population heuristic (genetic algo- Dear.9 Computational results were presented for three
rithm) for the problem. Other work relating to the static simulated scenarios involving 500 aircraft and one runway.
ALP is described in Beasley et al1 and, for reasons of space, Brinton10 presented a tree search approach for the ALP.
that description will not be repeated here. In his approach, the tree represents the sequence in which
In this paper, we deal with the dynamic, or on-line, ALP, aircraft should be landed. The dynamic ALP is dealt with via
where decisions about the landing times (and runways) for freezing the position of an aircraft in the landing sequence
aircraft must be made as time passes and as the operational and via costs associated with changes in the scheduled
environment changes (aircraft land, new aircraft appear, landing time. No detailed computational results were
etc). Current air traffic control practice3–5 for dealing with presented however. Venkatakrishnan et al11 observed
this problem is to schedule aircraft to land in a first-come, separation times adopted on landing at Logan Airport
first-served (FCFS) manner. Boston. Using these observed separation times they applied
the work of Psaraftis,12,13 which they modified in a heuristic
manner to take account of aircraft time windows, to see the
improvement that could result from better sequencing. They
Previous work
presented two approaches to the dynamic ALP. In both
There appears to have been relatively little work published approaches aircraft are frozen in the landing sequence once
with regard to the dynamic ALP. Andreussi et al6 referred to they are near to landing. The difference between their
the problem as the aircraft sequencing problem and approaches is that one leaves the aircraft time window
presented a paper concerned with developing a discrete- unchanged, while the other gradually reduces the size of the
event simulation model to evaluate different sequencing time window as an aircraft approaches landing. Computa-
strategies. Computational results were presented for a tional results were presented for six data sets involving up to
number of simulated scenarios involving three runways. 92 aircraft and two runways.
Dear and Sherif7,8 discussed both the static and dynamic Ciesielski and Scerri14,15 presented a genetic algorithm for
ALP and presented a heuristic algorithm for the single the problem. In their approach, landing times are allocated
runway dynamic ALP based upon constrained position by specifying a 30-s time slot. Their approach consists of
shifting. This involves finding, for a small set of aircraft, finding the best solution they can within 3 min of elapsed
the best possible positions for them in the landing queue time, updating the situation with respect to aircraft that have
subject to the constraint that no aircraft can be moved more landed/appeared and resolving a new problem. Unlike the
than a pre-specified number of positions away from the work presented in this paper, they include no link between
position it had in the landing queue based on FCFS, see also the previous set of landing time decisions and the new set.
JE Beasley et al—Displacement problem and scheduling 57
Computational results were presented for two data sets Similarly, if the assigned landing time Xi is less than the
involving 28 and 29 aircraft and two runways. Milan5 target time Ti, then when the displacement problem is
considered the problem of assigning priorities for landing to solved, aircraft i will not wish to move further away from its
aircraft in arrival batches (a batch comprising aircraft due to desired target time Ti and any such movement should be
arrive at approximately the same time). Priorities were based penalised by an additional cost. Again for the sake of
upon factors such as number of passengers, cost of passenger illustration we shall assume, as in Figure 1, that this extra
delays and proportion of transfer passengers. Once priorities cost is gi for each unit of time the aircraft is displaced
had been assigned the aircraft in a batch were landed in earlier than Xi, that is, that the displacement function is
priority order. Computational results were presented for one gi max[0, Xixi] if XioTi. Hence our displacement func-
example with 30 aircraft and one runway. tion is:
Carr et al16–18 presented papers concerned with modifying
the standard FCFS approach to allow individual airlines to Di ðX; xÞ ¼gi max½0; Xi xi if Xi oTi
express priorities with regard to the landing of their aircraft. hi max½0; xi Xi if Xi 4Ti
ð9Þ
Computational results were presented for a number of gi max½0; Xi xi þ hi max½0; xi Xi
simulated scenarios involving three runways. Bolender and if Xi ¼ Ti
Slater19 approached the dynamic ALP via queueing theory
and discrete-event simulation. They assumed that aircraft where for the case Xi ¼ Ti we ensure that deviations from
appear according to a Poisson process so that aircraft target (in either direction) are further penalised. This
interarrival times follow a negative exponential distribution. displacement function satisfies the restrictions (D(X, x)
Their analysis focuses on differing ways of allocating a newly non-negative, zero if xi ¼ Xi and a separable function)
appeared aircraft to a runway for landing (when multiple mentioned previously.
runways are present). Once allocated to a runway aircraft
land in FCFS order. Computational results were presented
Solving the displacement problem
for a number of problems involving up to three runways.
Wong20 discusses the algorithms underlying the CTAS In order to solve the displacement problem we adapted
(Center TRACON (terminal radar approach control) three solution approaches, one optimal1 and two heuristic,1,2
Automation System) system developed at the NASA Ames given previously in the literature for the static ALP. We
Research Center. For the dynamic ALP, an FCFS approach believe it to be a natural state of affairs that solution
is used with aircraft being frozen once they are close to approaches (both heuristic and optimal) developed for the
landing. static ALP can with suitable amendment be applied directly
to the dynamic ALP. One should reasonably expect that
solution approaches developed for the original static
Displacement function
decision problem have an important role to play in solving
In order to deal with the dynamic ALP as a displacement the displacement problem defined from the original static
problem, we need to define an appropriate displacement decision problem. The adaptations we made were as briefly
function. Considering Figure 1 suppose that aircraft i has outlined below.
been assigned a landing time Xi from the solution to the Beasley et al1 presented an optimal solution algorithm
original static ALP and further suppose that this time is later based upon linear programming (LP)-based tree search for
than its desired target time (ie Xi4Ti). When we come to the static ALP. Although the displacement function
solve the displacement problem the effect of the cost (Equation (9)) defined above for the dynamic ALP is
component (gi max[0, Tixi] þ hi max[0, xiTi]) associated nonlinear, by utilising the definition of the displacement
with aircraft i in the displacement problem objective function problem (Equations (1)–(4)) together with the definition of
will be to try and move the new landing time for aircraft i the static ALP (Equations (5)–(8)), it is possible to show that
(xi) closer to the desired target time Ti (ie to have xioXi). the dynamic ALP as formulated above can be transformed
This will be a desirable displacement from the current into a mixed-integer zero–one linear program. Hence, LP-
landing time Xi. based tree search can be directly applied in order to find the
It may be however that other factors (eg newly appeared optimal solution to each successive displacement problem.
aircraft that must be scheduled for landing) will mean that We refer to this algorithm as DALP-OPT.
aircraft i has its landing time further increased (ie xi4Xi). The heuristic solution algorithm presented in Beasley et al1
This will not be a desirable displacement and so should incur for the static ALP used two steps: firstly a simple
an extra cost, such as shown by the dotted displacement constructive step to decide the sequence in which aircraft
function line in Figure 1. For the sake of illustration we shall are to land (and on which runway), based upon sorting
assume, as Figure 1, that this extra cost is also hi for each aircraft into target time order; and secondly solving ALPF to
unit of time the aircraft is displaced later than Xi, that is, that decide the landing times for each aircraft. With respect to
the displacement function is hi max[0, xiXi] if Xi4Ti. adapting this heuristic for the solution of the displacement
58 Journal of the Operational Research Society Vol. 55, No. 1
problem we found that, computationally, directly applying current time had its landing time (and runway) frozen.
this heuristic to the displacement problem could lead to Set the landing times Xi ¼ N, i ¼ 1,y,P. Let:
infeasibility, that is, we could end up with a displacement
problem for which the heuristic could not find a feasible F0(t) represent the set of aircraft that have not yet
solution. We found that, for the particular test problems appeared by time t, that is, F0(t) ¼ [i|Ai4t,
we considered, this issue of infeasibility did not arise if i ¼ 1,y,P]
we also generated a landing sequence based upon first F1(t) represent the set of aircraft that have appeared by
sorting aircraft using their landing time as decided at the time t, but have not yet landed (or had their landing
solution to the previous displacement problem and then times frozen), that is, F1(t) ¼ [i|Aipt and Xi4t þ t*,
including any newly appeared aircraft in target time i ¼ 1,y,P]
order. Again given a landing sequence we solved ALPF F2(t) represent the set of aircraft that have appeared by
to decide landing times. In the computational results time t and have landed (or have had their landing
presented below we applied both of these heuristics and times frozen), that is, F2(t) ¼ [i|Aipt and Xipt þ t*,
took the best solution found. We refer to this algorithm as i ¼ 1,y,P]
DALP-H1. g be the iteration counter
In order to adapt the population heuristic (genetic Zdisp be the accumulated displacement cost
algorithm) presented by Beasley et al2 for the single runway
static ALP to the multiple runway dynamic ALP, we: (2) Set g ¼ 0 and Zdisp ¼ 0. Set the current time
extended the representation, together with the crossover t0 ¼ min [Ai|i ¼ 1,y,P] and solve the original static
and mutation operators, to include runway choice; ALP involving just those aircraft in F1(t0). The solution
incorporated maximum displacement by time window to this static ALP gives the initial landing times Xi
modification; 8iAF1(t0).
seeded the initial population with suitable individuals (3) If aircraft are still to appear (|F0(tg)|X1) then go to step
based upon the landing sequences given by sorting aircraft (4), otherwise (|F0(tg)| ¼ 0) all aircraft have appeared in
into earliest/target/latest time order; which case go to step (5).
applied the population heuristic after first using DALP- (4) Set g ¼ g þ 1. Advance the time to tg ¼ min[Ai|iAF0(tg1)]
H1 and seeding the initial population with the best and solve the displacement problem involving just those
solution found by DALP-H1 aircraft in F1(tg),F2(tg), where the aircraft in F2(tg) are
took the best solution as found by the population heuristic constrained to land at the landing times Xi (8iAF2(tg)),
and solved ALPF to decide landing times. and on the appropriate runways, as were obtained from
the previous displacement solution. Add the displace-
P
We refer to this algorithm as DALP-H2. ment cost component ðldisp i2F1 ðtg Þ\F1 ðtg1 Þ pi Di ðX; xÞ þ
lmax Dmax Þ of this solution to Zdisp and go to step (3).
(5) All aircraft in F1(tg) are now deemed to land at
Computational results the landing times (and on the appropriate runways) as
The algorithms DALP-H1, DALP-H2 and DALP-OPT for were obtained from the last displacement solution.
P
the dynamic ALP outlined in this paper were programmed in Compute Zsol ¼ Pi¼1 (gi max[0, TiXi] þ hi max[0,
FORTRAN and run on a Silicon Graphics Indigo work- XiTi]) which is the cost of the final solution in terms
station (R4000, 100 MHz, 64 MB main memory) for a of the cost function associated with the original (static)
number of test problems involving up to 500 aircraft and five ALP.
runways. This machine is approximately 25 times slower
than a 2.5 GHz Pentium pc. In order to solve the mixed-
integer zero–one formulation of the displacement problem Results
to optimality using LP-based tree search, and also to
We considered two sets of test problems. The first set,
solve ALPF, we used the CPLEX (version 6.5) software
involving up to 50 aircraft, were those previously considered
package.21
in Beasley et al1 for which optimal static solutions are
known. The second set we considered were larger, involving
Methodology between 100 and 500 aircraft, and were generated in the
following manner:
The methodology we adopted was as follows:
(1) Each aircraft i (i ¼ 1,y,P) had an appearance time Ai, aircraft appeared according to a negative exponential
the time at which it was first available to be assigned a distribution with a mean interarrival distance of 6.5
landing time. We also defined a freeze time t* such that nautical miles, converted into an earliest time by assuming
any aircraft assigned a landing time within t* of the a speed of 210 knots (nautical miles per hour)
JE Beasley et al—Displacement problem and scheduling 59
the appearance time for each aircraft was 10 min before above the static solution is 591.1% for FCFS
its earliest time; the target time was randomly generated (100(ZFCFSZstatic)/Zstatic) 55.1% for DALP-H1
between 1 and 10 min after the earliest time; the latest time (100(Zsol þ ZdispZstatic)/Zstatic), 50.9% for DALP-H2,
was 30 min after the earliest time but only 36.4% for DALP-OPT.
costs for appearing before/after target were real numbers Of the 39 larger problems shown in Tables 2 and 3, we
randomly generated from the interval [1,2] and a freeze have that the best solution value (Zsol þ Zdisp) is given by
time of 12 min was adopted DALP-OPT 37 times. This compares with 15 times for
aircraft were classified by four types: heavy, upper- DALP-H2 and eight times for DALP-H1 and indicates
medium, lower-medium and small, with the type being the value of resolving each successive displacement
randomly generated with probabilities of 0.4, 0.3, 0.2 and problem optimally rather than heuristically. Note here
0.1 respectively. Separation distances/times on landing that normally one expects an optimal algorithm (by its
were calculated as in2. very nature) to always produce a solution superior (or
equal) to that produced by a heuristic algorithm. Here
For the largest problem solved this corresponds to however for two of these 39 problems (problem 12 with
scheduling the landing of 500 aircraft over a 15-h time three runways in Tables 2 and 3), our heuristic DALP-H2
period. All of the test problems solved in this paper are produces a better solution than DALP-OPT. The reason
publicly available from OR-Library,22,23 see http:// for this is a generic one in that we are solving a succession
mscmga.ms.ic.ac.uk/jeb/orlib/airlandinfo.html. of displacement problems as time passes (aircraft land,
Table 1 gives the results for the smaller problems with new aircraft appear) and the values shown in Tables 2 and
lcost ¼ ldisp ¼ 1, lmax ¼ 0 and pi ¼ 1, Di ¼ N 8i. In that 3 are summary statistics of the overall effect of these
table we, in order to provide some insight into the quality solutions in cost terms. It can happen, as here, that solving
of our results, have given the solution value associated a displacement problem heuristically leads to decisions
with the optimal static approach (from Beasley et al1). that are better (in terms of aircraft yet to appear — which
This value provides a benchmark since (in terms of the are unknown) than the decisions made by solving the
original cost, Equation (5)) the best possible sequence of same displacement problem optimally. As such, the
decisions (landing times and runways) we can make, over a overall solution produced by successive applications of a
succession of displacement problem solutions, correspond to heuristic algorithm can be superior to the overall solution
the decisions (landing times and runways) arrived at by produced by successive applications of an optimal
solving, just once, the static ALP involving all aircraft. We algorithm.
also show in Table 1 the solution value found by taking It is clear that the benefit gained by DALP-H2,
the best solution from two FCFS approaches: schedule compared to DALP-H1, is more marked for larger
each aircraft as early as possible; and schedule each aircraft problems. DALP-H2 improves upon the DALP-H1
at its target time (if that is feasible) but as early as possible solution (Zsol þ Zdisp) for only three of the 25 problems
if the target time is not feasible. Summarising then we in Table 1, but for 31 of the 39 problems in Tables 2
have that Table 1 shows, for each problem: the number of and 3.
aircraft; the number of runways; the optimal static
solution value (Zstatic); the FCFS solution value (ZFCFS);
In order to investigate problems where DALP-OPT becomes
the solution values (Zsol, Zdisp) and the total time taken
computationally ineffective we show in Table 4 the results
when each successive displacement problem is solved
heuristically/optimally. for the same problems as in Table 2 but with the freeze time
Table 2 shows similar information as Table 1 for the reduced to zero. Reducing the freeze time increases the
larger problems. Table 3 also shows results for these number of aircraft that are available to have their landing
problems but where, compared to Table 2, we have increased times rescheduled and hence increases the size of the
ldisp to 2, lmax to 5, and have set Di ¼ 10 8i. This corresponds displacement problem that has to be solved optimally by
to a scenario in which we are trying to further discourage DALP-OPT. It is clear from Table 4 that while for some
displacement and also explicitly limit displacement. We only problems DALP-OPT is still computationally effective, there
show in Table 3 those test problems that exhibited some are a number of problems (in particular those that
displacement in Table 2. Examining Tables 1–3 we would terminated without a solution due to time limit considera-
make the following observations: tions) for which DALP-OPT is computationally ineffective.
Note here, however, that both DALP-H1 and DALP-H2 are
Excluding from Table 1 those problems for which computationally effective for all of the problems shown in
Zstatic ¼ 0, the average percentage cost increase over and Table 4.
60 Journal of the Operational Research Society Vol. 55, No. 1
Table 1 Computational results: small problems
Problem Number of Number of Optimal FCFS Heuristic solution to each successive displacement problem Optimal solution to
number aircraft runways static solution each successive
solution ZFCFS displacement
Zstatic problem DALP-OPT
DALP-H1 DALP-H2
Zsol Zdisp Total time Zsol Zdisp Total time Zsol Zdisp Total time
(min) (min) (min)
1 10 1 700 1790 740 260 0.1 Same 0.1 740 260 0.1
2 90 120 120 0 0.1 Same 0.1 90 30 0.1
3 0 0 0 0 0.1 Same 0.1 0 0 0.1
2 15 1 1480 2610 1870 110 0.1 Same 0.2 1730 250 0.5
2 210 210 210 0 0.2 Same 0.2 210 0 0.3
3 0 0 0 0 0.1 Same 0.1 0 0 0.1
3 20 1 820 2930 1440 290 0.2 Same 0.4 940 230 0.5
2 60 60 60 0 0.2 Same 0.3 60 0 0.3
3 0 0 0 0 0.1 Same 0.1 0 0 0.1
4 20 1 2520 6290 2670 960 0.2 Same 0.3 2700 420 0.9
2 640 1560 680 170 0.2 Same 0.3 680 80 2.0
3 130 330 130 10 0.2 Same 0.3 130 10 0.6
4 0 60 0 0 0.1 Same 0.1 0 0 0.1
5 20 1 3100 8370 6130 490 0.3 6130 250 0.4 3810 630 1.3
2 650 1440 1070 60 0.2 1050 120 0.3 680 170 4.0
3 170 240 240 0 0.2 Same 0.3 240 0 0.9
4 0 0 0 0 0.1 Same 0.1 0 0 0.1
6 30 1 24 442 24 442 24 442 0 0.4 Same 0.6 24 442 0 0.6
2 554 882 882 0 0.4 Same 0.4 809 98 0.5
3 0 0 0 0 0.2 Same 0.2 0 0 0.2
7 44 1 1550 1550 3974 0 1.0 Same 1.4 3974 0 1.5
2 0 0 0 0 0.6 Same 0.6 0 0 0.6
8 50 1 1950 26835 2915 735 1.3 2710 430 2.0 2000 455 3.1
2 135 10140 255 15 1.1 135 75 1.2 135 75 1.2
3 0 4825 0 0 0.7 Same 0.7 0 0 0.7
Note: ‘Same’ means that the solution values for both Zsol and Zdisp for DALP-H2 were identical to those for DALP-H1.
Table 2 Computational results: large problems
Problem Number of Number of FCFS Heuristic solution to each successive displacement problem Optimal solution to each successive
number aircraft runways solution displacement problem DALP-OPT
ZFCFS
DALP-H1 DALP-H2
Zsol Zdisp Total time Zsol Zdisp Total time Zsol Zdisp Total time
(min) (min) (min)
9 100 1 36 839.36 13 554.77 1213.12 3.7 12 553.87 1041.40 4.7 7848.42 1171.12 5.7
2 10 661.92 578.71 57.94 3.0 577.57 33.12 2.8 573.25 33.12 3.4
3 4142.07 88.72 33.28 2.6 Same 2.6 88.72 33.28 2.6
4 1518.04 0 0 2.3 Same 2.3 0 0 2.3
10 150 1 54 113.54 31 945.03 1602.79 7.6 31 034.04 1220.52 9.7 17 726.06 2164.97 13.5
2 13 318.09 1903.00 428.36 6.9 1417.82 307.82 6.8 1372.21 53.70 8.7
3 5290.35 219.44 34.68 5.7 216.25 34.68 5.4 246.15 1.14 6.0
4 3203.35 47.20 0 5.6 34.22 0 5.1 34.22 0 5.1
5 770.90 0 0 4.9 Same 4.9 0 0 4.9
11 200 1 66 427.28 27 417.23 2714.15 16.6 23 963.48 2150.87 18.5 19 327.45 1954.41 22.6
2 17 381.29 1671.13 455.31 12.3 1692.97 245.84 12.1 1683.75 91.32 15.9
3 5901.27 347.37 0 10.2 333.53 0 9.4 333.53 0 13.2
4 2040.41 69.66 0 8.9 Same 8.9 69.66 0 9.3
5 516.84 0 0 8.6 Same 8.6 0 0 9.1
12 250 1 81 916.40 34 246.39 2599.64 22.3 31 439.77 1792.73 28.2 25 049.24 1941.88 43.2
2 22 790.99 2848.87 718.05 18.6 2497.06 545.68 19.3 2204.96 155.31 43.2
3 8883.11 265.24 44.10 15.8 252.52 44.10 15.7 430.50 37.74 17.1
4 3261.29 2.86 0 14.1 Same 14.1 2.86 0 14.1
5 1053.63 0 0 13.5 Same 13.5 0 0 13.5
DALP-H1 DALP-H2
Zsol Zdisp Total time Zsol Zdisp Total time Zsol Zdisp Total time
(min) (min) (min)
Zsol Zdisp Total time Zsol Zdisp Total time Zsol Zdisp Total time
(min) (min) (min)
9 100 1 8185.25 3762.06 4.1 7800.93 2947.55 5.5 6219.54 1371.35 77.1
2 510.40 205.44 3.8 480.00 117.12 4.6 480.58 83.40 9.7
3 88.72 33.28 3.2 Same 3.2 88.72 33.28 4.0
4 0 0 2.3 Same 2.3 0 0 2.4
10 150 1 23 569.43 4105.49 9.0 23 137.89 2885.86 12.0 (781.2:59)
2 1361.51 418.55 7.7 1385.02 373.05 8.9 1317.69 81.00 119.3
3 218.29 65.97 6.1 216.25 34.68 6.6 246.15 1.14 24.7
4 47.20 0 5.1 34.22 0 5.1 34.22 0 6.0
5 0 0 4.9 Same 4.9 0 0 5.0
11 200 1 16 474.79 5064.40 16.9 16 865.47 4843.90 21.5 14 152.40 2204.73 134.3
2 1655.53 493.03 14.4 1654.30 484.10 17.1 1587.39 109.06 177.5
3 347.37 0 10.5 333.53 0 11.0 333.53 0 485.6
4 69.66 0 9.3 Same 9.5 69.66 0 31.4
5 0 0 8.8 Same 8.8 0 0 8.9
12 250 1 25 547.94 4842.58 27.2 25 750.13 3922.92 39.9 (631.5:38)
2 1934.68 778.28 21.8 1921.38 790.78 26.0 (833.4:37)
3 265.24 44.10 15.2 Same 15.6 252.52 44.10 632.2
4 2.86 0 14.0 Same 14.0 2.86 0 15.6
5 0 0 13.7 Same 13.7 0 0 13.9
Acknowledgements—JE Beasley acknowledges the financial support of 11 Venkatakrishnan CS, Barnett A and Odoni AR (1993). Land-
the CSIRO Australia. We also acknowledge comments made on earlier ings at Logan Airport: describing and increasing airport
versions of this paper by referees. capacity. Transport Sci 27: 211–227.
12 Psaraftis HN (1978). A dynamic programming approach to the
aircraft sequencing problem, Report R78-4 Flight Transporta-
tion Laboratory, MIT, Cambridge MA, USA.
References 13 Psaraftis HN (1980). A dynamic programming approach for
sequencing groups of identical jobs. Opns Res 28: 1347–1359.
1 Beasley JE, Krishnamoorthy M, Sharaiha YM and Abramson 14 Ciesielski V and Scerri P (1997). An anytime algorithm for
D (2000). Scheduling aircraft landings — the static case. scheduling of aircraft landing times using genetic algorithms.
Transport Sci 34: 180–197. Aust J Intell Inf Process Systems 4: 206–213.
2 Beasley JE, Sonander J and Havelock P (2001). Scheduling 15 Ciesielski V and Scerri P (1998). Real time genetic scheduling of
aircraft landings at London Heathrow using a population aircraft landing times. In: Fogel D (ed) Proceedings of the 1998
heuristic. J Opl Res Soc 52: 483–493. IEEE International Conference on Evolutionary Computation
3 Odoni AR, Rousseau J-M and Wilson NHM (1994). Models in (ICEC98). IEEE, NY, USA, pp 360–364.
urban and air transportation. In: Pollock SM, Rothkopf MH 16 Carr GC, Erzberger H and Neuman F (1998). Airline arrival
and Barnett A (eds) Operations Research and the Public Sector: prioritization in sequencing and scheduling. Paper presented at the
Handbooks in Operations Research and Management Science, Vol second USA/Europe Air Traffic Management R&D Seminar,
6. North-Holland: Amsterdam, The Netherlands, pp 107–150. Orlando. Available from. http://www.ctas.arc.nasa.gov/
4 Erzberger H (1995). Design principles and algorithms for publications.
automated air traffic management: In: Knowledge-based Func- 17 Carr GC, Erzberger H and Neuman F (1999). Delay exchanges
tions in Aerospace Systems. AGARD Lecture Series no. 200. in arrival sequencing and scheduling. J Aircraft 36: 785–791.
NATO Neuilly-Sur-Seine, France, 7:1–7:31. 18 Carr GC, Erzberger H and Neuman F (2000). Fast-time study
5 Milan J (1997). The flow management problem in air traffic of airline-influenced arrival sequencing and scheduling. J
control: a model of assigning priorities for landings at a Guidance Control Dyn 23: 526–531.
congested airport. Transport Plann Technol 20: 131–162. 19 Bolender MA and Slater GL (2000). Evaluation of scheduling
6 Andreussi A, Bianco L and Ricciardelli S (1981). A simulation methods for multiple runways. J Aircraft 37: 410–416.
model for aircraft sequencing in the near terminal area. Eur J 20 Wong GL (2000). The dynamic planner: the sequencer, scheduler,
Opl Res 8: 345–354. and runway allocator for air traffic control automation, Report
7 Dear RG and Sherif YS (1989). The dynamic scheduling of NASA/TM-2000-209586, NASA Ames Research Center, Mof-
aircraft in high density terminal areas. Microelectron Reliab 29: fett Field, CA, USA, Available from. http://www.ctas.arc.na-
743–749. sa.gov/publications.
8 Dear RG and Sherif YS (1991). An algorithm for computer 21 ILOG Inc (1999). ILOG CPLEX 6.5 User’s Manual. ILOG Inc.:
assisted sequencing and scheduling of terminal area operations. Mountain View, CA, USA.
Transport Res A 25A: 129–139. 22 Beasley JE (1990). OR-Library: distributing test problems by
9 Dear RG (1976). The dynamic scheduling of aircraft in the near electronic mail. J Opl Res Soc 41: 1069–1072.
terminal area, Report R76-9 Flight Transportation Laboratory, 23 Beasley JE (1996). Obtaining test problems via Internet.
MIT, Cambridge, MA, USA. J Global Optim 8: 429–433.
10 Brinton CR (1992). An implicit enumeration algorithm for
arrival aircraft scheduling. In: Proceedings of the 11th IEEE/
AIAA Digital Avionics Systems Conference Seattle, WA. IEEE, Received January 2003;
NY, USA, pp 268–274. accepted October 2003 after two revisions