Local Search and Constraint Programming for the Post Enrolmentbased Course Timetabling Problem


 Amos Porter
 6 years ago
 Views:
Transcription
1 Local Search and Constraint Programming for the Post Enrolmentbased Course Timetabling Problem Hadrien Cambazard, Emmanuel Hebrard, Barry O Sullivan and Alexandre Papadopoulos Cork Constraint Computation Centre Department of Computer Science, University College Cork, Ireland {h.cambazard e.hebrard b.osullivan Abstract. We present a study of the university postenrolment timetabling problem, proposed as Track 2 of the 2007 International Timetabling Competition. We approach the problem using several techniques, particularly local search, constraint programming techniques and hybrids of these in the form of a large neighbourhood search scheme. Our local search approach won the competition. Our best constraint programming approach uses an original problem decomposition. Incorporating this into a large neighbourhood search scheme seems promising. 1 Introduction Timetabling problems have a wide range of applications in education, sport, manpower planning, and logistics. A diverse variety of university timetabling problems exist, but three main categories have been identified [5, 9, 26]: school, examination and course timetabling. The Post Enrolment University Course Timetabling Problem [17] occurs in an educational context whereby a set of events (lectures) have to be scheduled in timeslots and located in appropriate rooms. The problem tackled in this paper was proposed as part of the 2007 International Timetabling Competition organised by PATAT (Track 2) 1. The problem was also used in the 2003 competition without two specific hard constraints introduced in 2007, which are discussed in Section 2. These new constraints were introduced in the 2007 competition in order to make the search for feasible timetables difficult. In 2003 finding feasible timetables was relatively easy and all algorithms, therefore, focused on optimising the soft constraints. According to the organisers [17], the two constraints have been added to move the competition closer to realworld timetabling where finding feasible timetables can be a very challenging task. This context seemed a good opportunity to investigate Constraint Programming (CP) techniques, and compare them with the strong local search baseline developed during the 2003 challenge. Our main contribution in this paper is a comprehensive study of the problem using a wide range of techniques highlighting both pitfalls and positive results. Our main technical novelty lies in the analysis of complete approaches with This work was supported by Science Foundation Ireland (Grant Number 05/IN/I886). 1
2 original CP models and lower bounds for the costs associated to the soft constraints, including algorithms to maintain them. We also present an original local search approach that can deal with the hardness of feasibility; this was ranked first out of thirteen teams in Track 2 of the 2007 International Timetabling Competition. Finally, a promising large neighbourhood search (LNS) scheme [27] is proposed, which contrasts with all previous published local search work on this problem [1, 6, 10, 15, 24]. 2 Problem Description The post enrolmentbased course timetabling problem consists of a set of n events, E, to be scheduled in 45 timeslots {1,..., 45} (5 days of 9 hours each) using a set of m rooms, R. Each room is characterised by its seating capacity, which we will refer to as its size, and a set of features defining the set of services available in each room. Each event needs a room whose size is larger than the number of students attending the event, and it must be placed in a room with the required features. Additionally, a set of precedence requirements state that some events must occur before others. We are also given a set S of students and the set of events each must attend. Each event must be assigned to a room in a timeslot while obeying a set of constraints. The constraints of the problem are partitioned into two sets: the hard constraints define the requirements of a feasible timetable, while the soft constraints define an optimal timetable. The hard constraints are the following: 1. No student can attend more than one event at the same time. 2. In each case the room has to be big enough for all the attending students and satisfy all of the features required by the event. 3. Only one event is put into each room in each timeslot. 4. An event can only be assigned to its predefined available timeslots. 5. When specified, events have to occur in the correct order in the week. Because feasibility can be very difficult to achieve, the organisers of the competition have introduced the notion of distance to feasibility to be able to discriminate entries that do not find any feasible solution. We will ignore this point in our study and consider all infeasible solutions as mere failures 2. The quality of a feasible timetable is evaluated in terms of the soft constraints. A feasible solution is penalised equally if a student: 1. attends an event in the last timeslot of the day ({9, 18, 27, 36, 45}); 2. attends more than two events in a row on a given day; one penalty is counted for each event attended consecutively after the first two; 3. attends exactly one event during a day. The problem defined by the hard constraints only can be seen as a constrained listcolouring in a graph where a node is an event and an edge is added between two events that must go to different timeslots. This graph is primarily made of many large overlapping cliques, referred to as student cliques, defined by the set of events chosen by each student. It can also be worthwhile to notice that two events that share a unique 2 This notion is indeed relevant for the competition but not for the problem itself.
3 Table 1. Some statistics about the colouring graph structure in the first eight instances. inst student cliques room cliques final cliques density min max avg number min max avg number min max avg basic full possible room, due to their size and features, have to be assigned to different timeslots. The cliques relying on those edges are referred to as room cliques. At last, precedences also imply differences and can be added in the graph. Table 1 gives some details about the size and number of cliques found in the colouring graph because both our LS and CP approaches will try to take advantage of them. It also shows the density of the basic graph, i.e. the original graph of student choices including the precedence edges, and the full graph, i.e. the same graph augmented with rooms. The final cliques of Table 1 are obtained by the following process: the neighbourhood of a clique c, i.e. the set of nodes connected to all the nodes of c (but not necessarily with each other) can intersect another clique, and the corresponding intersection can, therefore, be used to extend c. The final cliques are obtained by applying such a process iteratively from the student/room cliques until a fixed point is reached. The density of the full graph is not much bigger than for the basic graph, but the added edges can significantly improve the maximum and average size of the cliques. 3 A Local Search Approach Our local search baseline is strongly based on the work achieved during the 2003 competition and the improved results published later in [6, 15, 24]. There are, however, some differences to consider due to the increased difficulty of finding feasible timetables in the 2007 competition instances. Similar to most approaches of 2003, our local search is performed in two steps: we first try to identify a feasible solution, and then try to reduce the cost of violating the soft constraints. The originality of our local search lies mainly in finding feasible timetables. We describe both steps in more detail below. 3.1 Finding Feasible Solutions The search for feasible solutions is performed by considering a unit cost per hard constraint violation: an infeasible timeslot or room for an event, two events sharing a student in the same timeslot, two events violating a precedence between them. Representation of the solution. The position of an event is defined by a given timeslot and room. The solution is represented by the position of each event as opposed to the solution representation described in [24], which ignores the rooms and maintains the room violations by solving a matching problem per timeslot. Knowing if a set of events can fit in a given timeslot with respect to room availability and capacity is a
4 bipartite matching problem (events to rooms). For efficiency reasons, the lists of events per timeslot as well as the list of all free positions in the timetable (positions where no event is currently assigned) are added to the representation. Neighbourhood. The neighbourhood can be seen as a composite neighbourhood structure [1, 10] defined in terms of the following moves: 1. T re: translates an event to a free position of the timetable. 2. SwE: swaps two events by interchanging their position in the timetable. 3. SwT : swaps two timeslots t i and t j, i.e. translates all events currently placed in t i to t j and all events in t j to t i. 4. Ma (Matching): reassigns the events within a given timeslot to minimise the number of room conflicts; to allow violations, a maximum matching is solved. 5. T re+ma: translates an event to a timeslot and evaluates if this does not violate the room constraints by checking the corresponding matching problem; if the matching is infeasible, the move is rejected. 6. Hu (Hungarian): picks a set of events {e 1,..., e k } assigned in different timeslots (k 45) that do not have precedences defined between them, and reassigns them optimally by solving an assignment problem with the Hungarian algorithm [16]. The violation of the hard constraints for placing each event in each timeslot is known as it does not depend on the other removed events, since they do not share precedences and only a single event is removed per timeslot. We solve 45 k maximum matching problems to evaluate the cost due to the room capacities of placing each event in each timeslot. As the number of such moves is exponential, the size of the neighbourhood is restricted to K sets (K = 20 in practice), including conflicting events (involved in hard constraints violations) and completed randomly. T re and SwT are always considered in the neighbourhood. The remaining moves are ranked in terms of their time complexities and included in the neighbourhood at a given iteration depending on a probability related to their complexity. More specifically, the probabilities are set to p(hu) = p(swt ) = 5 10, p(ma) = , p(t re + Ma) = Thus, time consuming moves are performed less frequently than faster ones. The 2 set of moves considered at each iteration, therefore, varies and the order of exploration amongst them is chosen randomly. However, for each move the exploration is performed deterministically from the last point where it was left (similar to [15]). Search. Improving and sideways moves, which keep the current violation cost constant, are always accepted and no emphasis is put on conflicting events, except by move Hu. We believe that moves T re and SwE are very important in our approach. They can be performed very quickly and, therefore, provide a diversification mechanism as the search is not guided by conflicting events. This also explains why we choose a solution representation that includes the room information explicitly, since this is mandatory for T re and SwE. A simple tabu list of size k = 10 prevents cycling by forbidding an event being considered in a timeslot it was assigned in the last k iterations; this is similar to [6] and classic in graph colouring [11]. Finally, a pure random configuration is used to start as we found no significant benefits to starting from a greedy solution.
5 Intensification. As mentioned previously, the problem defined by the hard constraints can be seen as a constrained listcolouring problem in which the graph is made of many overlapping large cliques (see Table 1). The intensification step tries to take advantage of the presence of such large cliques by iteratively applying move Hu on each clique containing at least one conflicting event. All events of the clique have to be in different timeslots and define an assignment problem in the current timetable. This intensification is applied every nonimproving iterations. All cliques containing a conflicting event are considered, and simplified to ignore any precedences amongst the events inside the clique. This step is applied on all the final cliques of Table 1. Results. Table 2 compares a local search LS1 with a neighbourhood based only on the moves {T re, SwE, Ma, T re +Ma} with the full scheme, LS2, described previously (involving a richer and randomised neighbourhood as well as the intensification step). All instances except instances 1,2,9 and 10 seem quite easy from the point of view of feasibility, and the efficiency of our improvements can be mainly seen on instances 2,9 and 10, which are much more challenging. The diversification given by the randomised neighbourhood (by favouring T re and SwE), and the intensification given by Hu and its systematic use on the cliques, is beneficial for the hard instances. Table 2 shows the percentage of feasible solutions found over 100 runs with different seeds within the time limit 3 and the average time required by LS1 and LS2, which is computed only on runs that have found a solution. Instance 10 is the only instance that remains really open for feasibility as all others are solved more than 94% of the time. Table 2. Percentage of solutions found, with average time, using a simple Local search (LS1) and our improved scheme (LS2). Instances LS1 % solved avg time (s) LS2 % solved avg time (s) Finding Good Solutions Once a feasible solution has been found, another local search optimises its soft cost. Representation of the Solution. We extend the previous representation by adding the student view. The timetable of each student (needed for cost 2) is kept as a three dimensional matrix of size S 5 9 where each entry is equal to the event attended by the student at the corresponding day and timeslot (if there is one, and set to 1 otherwise). Moreover, the number of events attended by each student, each day, is stored for cost 3. Neighbourhood. The only move used in this phase is T re +Ma. Moreover, the moves considered are only those preserving feasibility. We note that this is a severe disadvantage for the search due to the tightness of the hard constraints. The main motivation for the LNS approach of Section 6 is to compensate for this disadvantage. 3 These experiments were run on a MacBook (2 GHz Intel Core Duo, 2 GB 667 Mhz DDR2) with a time limit of 420s given by the benchmarking system of the competition.
6 Search. The tabu search appears inefficient for the soft cost and better results are obtained using simulated annealing (SA) [14]. This seems to match the experience of [6, 15] and the study made in [24]. Improving and sideways moves are always performed and degrading ones are accepted with a probability depending on their cost variation : P acceptance (, τ) = e τ where the parameter τ, the temperature, controls the acceptance probability and is decreased over time. The temperature is cooled at each step using a standard geometric cooling τ n+1 = 0.95 τ n. Two parameters are needed to define the cooling: the initial temperature τ 0, and the length of a temperature step, L, i.e. the number of iterations performed at each temperature level. As the time demand varies a lot from one instance to the other, we try to predict the speed of our soft solver during an initialisation phase by running the SA at a temperature of 1 for iterations and set τ 0 and L in the following way. Firstly, τ 0 is set to the average of the cost variation observed during the initialisation; then, based on the time needed to perform the initialisation, we get an estimation of the number of iterations that will be performed in the remaining time, I. By setting a final temperature to τ f = 0.2, we also know the number of temperature step, nbsteps, needed to go from τ 0 to τ f and I nbsteps therefore L is set to L =. A reheating is performed if the neighbourhood is scanned without accepting any moves. This can happen if the number of feasible moves is limited and the SA is more likely to reject all choices as the temperature decreases. 3.3 A Synthesis of the Local Search Approach We conclude the presentation of the local search approach by showing the behaviour of the search at the two stages, i.e. feasibility and optimisation on the plots of Figure violation cost temperature cost 10 cost temperature number of iterations (a) Tabu search  feasibility stage e e+06 2e e+06 3e+06 number of iterations (b) Simulated Annealing  optimisation stage Fig. 1. Evolution of the violation cost per iteration for the two stages of the local search approach. Both plots show the evolution of the costs at each iteration. The cooling is also indicated for the SA. The search for feasibility proceeds by moving over a large plateau of configurations of equivalent violation cost, i.e. the cost is never degraded in practice. Sideways moves appear to be very frequent for feasibility. Therefore, the search can stay for a long time on the same plateau as it does not focus on conflicting events and accept
7 any sideways step; this is why favouring moves T re and SwE brings diversification over the plateau. Therefore, the role of the intensification step is important. Sideways moves are less likely to occur at the optimisation stage and one can see the effect of the cooling by observing that the cost variation is decreasing while the best known cost is converging toward its final value. The choice of the different metaheuristics for feasibility and optimisation, with their resulting behaviours, is also motivated by the fact that, in the first case, we try to get a feasible solution as soon as possible whereas in the second case we aim for the best possible solution within a given timelimit. 4 Constraint Programming Models for Feasibility This timetabling problem was tackled by a number of local search techniques [1, 6, 10, 15, 24]. We are not aware, however, of any complete approach. We considered several CP models, none of which were able to match the efficiency of local search. However, as we shall see in Section 6, the CP approach can still be valuable to provide complex neighbourhoods within the SA algorithm. We present here the most promising CP model as well as two less successful ones and give some insights into their inefficiency. 4.1 Basic Model For an event i we introduce two variables eventt ime i {1,..., 45} and eventroom i {1,..., m}, for the timeslot and room associated to event i, respectively. Let R i be the set of rooms that can accommodate event i, T i be the set of timeslots available for event i, student(i) be the set of students attending event i and, finally, let prec be the set of the pairs of ordered events. We define the first model as follows: Model 1 i, j n s.t. student(i) student(j) eventt ime i eventt ime j (1) i n eventroom i R i (2) i, j n (eventt ime i eventt ime j) (eventroom i eventroom j) (3) i n eventt ime i T i (4) (i, j) prec eventt ime i < eventt ime j (5) In this viewpoint, the constraints (1), (4) and (5) correspond to a list colouring problem with precedences on the variables eventt ime. Constraints (2) and (3) enforce that events be allocated to suitable rooms, and that within a given timeslot, every event be put into a different room, respectively. They correspond to a set of matching problems, one for each timeslot, conditioned by the result of the above colouring problem. An important observation is that these two aspects are relatively disconnected. Indeed, as long as an event is not committed to a given timeslot, we do not know in which matching it will participate because of the disjunctions (3). If an early decision on the colouring part prevents a consistent room allocation, it will not be discovered until very late in the search, leading to a trashing behaviour where large unsatisfiable subtrees are explored again and again. We explored two ways of resolving this issue. First, we modelled the relation between the room allocation (matching) and timeslot allocation (colouring) using a global constraint [2] to achieve stronger inference between these two aspects
8 and detect mistakes earlier. We describe this model in Section 4.2. The second solution was to separate the solving of the colouring and the matchings, so that we explore more diverse colourings, and hopefully avoid trashing. We describe this model in Section Matching Constraint Knowing if a set of events can fit in a given timeslot with respect to room availability and capacity is a bipartite matching problem (events to rooms). The objective is to remove the eventroom variables from the search space. In other words, we will make sure that constraint propagation alone ensures that an assignment of all events can be extended to a matching for each timeslot. As a result, we solve a colouring problem where we only assign events to timeslots subject to timeslot availability, precedences and such that the remaining matching subproblems are backtrackfree. The room allocation subproblem can be represented in a bipartite graph G = (V 1, V 2, E) where V 1 = {1,..., n} is the set of events, and V 2 = { 1, 1,..., 45, m } is the set of all pairs timeslot, room. An edge (i, j, k ) is present iff event i can be assigned to timeslot j in room k. A maximal matching of G thus represents an assignment of events to rooms satisfying constraints (2) and (3). We introduce n variables to link this matching with the colouring. event i { 1, 1,..., 45, m } denotes the timeslot and room, represented by a pair, to which event i is assigned. An ALLDIFF({event 1,..., event n }) [22] makes sure that the graph G admits a matching of cardinality n. Notice that during search, arcconsistency is achieved for all matching problems at once, giving stronger inference than considering matchings independently. Notice that we could post other constraints directly on variables events. However, this can be done more easily on the eventt ime variables. They also provide a naturally good branching scheme, since rooms have been factored out of the search space. We thus define a second model, where we substitute the variable eventroom i with event i and channel it to eventt ime i using a simple binary constraint. Model 2 i, j n s.t student(i) student(j) eventt ime i eventt ime j (1) i n eventt ime i T i (4) (i, j) prec eventt ime i < eventt ime j (5) i n event i R i T i (6) i n eventt ime i = event i[0] (7) ALLDIFF({event 1,..., event n}) (8) Constraint (7) channels the variables eventt ime to event by projecting on the first element of the pair. Notice that since arc consistency is achieved in polynomial time on the ALLDIFF constraint, an assignment of eventt ime satisfying Model 2 can always be extended to event in a backtrackfree manner. 4.3 Alternate Colourings and Matchings Constraint (8) of Model 2 is very costly to maintain. Therefore, we consider a decomposition similar to a logicbased Benders decomposition scheme [12]. We delay the resolution of the matchings once a colouring has been found. If the matching is infeasible, we seek another solution for the colouring subproblem, and iterate in this way until
9 a full solution is found. Clearly, solving the colouring part alone allows for a far more optimised and sleeker model, however, reaching a fixed point might not be easy. We first describe the lighter model restricted to the colouring and precedence constraints, and where the room allocation constraints are relaxed to a simpler cardinality constraint. Then, we show how Benders cuts can be inferred when failing to solve a matching in order to tighten the colouring subproblem. The eventroom variables are ignored as in the previous model and a single global cardinality constraint (GCC) [23] is added to ensure that every timeslot is used at most m times. This constraint eliminates trivially infeasible matchings where the number of events assigned to a timeslot is greater than the number of rooms. Model 3 i, j n s.t. student(i) student(j) eventt ime i eventt ime j (1) i n eventt ime i T i (4) (i, j) prec eventt ime i < eventt ime j (5) i n GCC({eventT ime i i n}, [[0..r],..., [0..r]]) (9) A solution of this model is not guaranteed to be a feasible solution of the original problem. Indeed, a matching problem can be inconsistent once the colouring is fixed. We, thus, iteratively solve the colouring part until we find a feasible room allocation, as depicted in Algorithm 1. If a matching problem fails, a minimal conflict corresponds to a set of events that cannot be assigned together in any timeslot, whilst forming an independent subgraph of the colouring graph. We use an algorithm for finding minimal conflicts [8] to extract such a set of events (line 3). In order to rule out this conflicting assignment in future resolutions of the colouring subproblem, we post a NOTALLEQUAL constraint to the model (line 4). The constraint NOTALLEQUAL(x 1,... x k ) ensures that there exists i, j [1..k] such that x i x k. It acts as a Benders cut and prevents the same assignment from being met again. Observe that since we extract minimal sets of conflicting events [4, 13], entire classes of assignments that would fail for the same reason are ruled out. Notice also that although this constraint is inferred from a particular timeslot, it holds for every timeslot. Algorithm 1: Decomposition 1 repeat 2 solve Model 3; matched true; foreach 1 j 45 do G (V 1 = {i eventt ime i = j}, V 2 = i V 1 R i, E = {(i, k) i V 1, k R i}); if cannot find a matching of G then matched false; 3 cut Extractminconflict(G); 4 add NOTALLEQUAL(eventT ime k k cut) to Model 3; until matched; We explored further improvements of this model based on the analysis of the colouring graph described in Section 2. Conflicts between events are organised into large cliques, one for each student and even larger cliques can be inferred by taking room
10 conflicts into account. This information can be used to obtain stronger filtering from the model. One possibility is to replace the constraints (1) by ALLDIFF constraints. Each of the aforementioned final cliques implies an ALLDIFF constraint between a set of eventt ime i variables. In this manner, all the binary differences (1) are covered by at least one clique and can thus be removed. We can expect to achieve a stronger level of propagation as a result. On the other hand, ALLDIFF can be expensive to maintain. We can therefore choose to keep, amongst the final cliques, only the cliques obtained from a room clique, as a tradeoff between the efficiency of binary differences and the additional reasoning brought by the cliques, as they are big and they contain additional conflicts. This leads to two variations of Model 3 that we assess empirically below. 4.4 Experimental Results We ran Model 2, Model 3, Model 3cliques (Model 3 including all implied ALLDIFF constraints) and Model 3rooms (Model 3 including only the ALLDIFF constraints standing for room cliques). In Table 3, we give the number of iterations of Algorithm 1 (Decompositon), that is, the number of feasible colourings that were required to find a complete solution. This number is always 1 for Model 2. We also give the cumulative CPU time and number of nodes explored on solved instances. Notice that no model could solve instances 1, 2, 9, 10, 13 or 14 within the time cutoff of 420 seconds, corresponding to the 10 minutes cutoff of the competition on an Apple MacBook. Model 2 does not need to solve several colouring problems, however, the overhead due to the extra variables (event) and to the large ALLDIFF constraint, is too large. In fact the search tree explored by Model 2 is several orders of magnitude smaller than that explored by Model 3. We also observe that in most cases, the ALLDIFF constraints on events sharing the same unique suitable room reduces dramatically the number of iterations required to solve the problems. On the other hand, using ALLDIFF constraints for representing the colouring problem seems to be slightly detrimental. The best combination seems to be Model 3 using ALLDIFF only for rooms. We believe that the main reason for Model 3 to dominate Model 2 is that the difficult part of the problem lies primarily in the colouring for these instances. The very low number of colouring subproblems solved when adding the implied ALLDIFF constraints provides further evidence of this. Any given colouring satisfying the implied ALLDIFF constraints is very likely to be extensible to a feasible matching. We also observed (but this is not apparent in the tables) that the extra GCC constraint used to approximate the matching part was almost unnecessary in most cases. That is, even without this constraint, the number of iterations to reach a complete solution remains relatively small. Notice, however, that this last observation does not stand for instances 1, 2, 9 and 10, which happen to be the hardest. Next we compare three heuristics all using the best model: Model 3 (room). We used minimum domain over future degree (dom/deg) and impact [21] as benchmarks, since they are both good general purpose heuristics. The former was used successfully on listcolouring problems in the past, whilst the latter proved to be the best in our experiments. The third heuristic, contention, is based on computing the contention of events for a given timeslot. In scheduling, resource contention has been used as heuristic with success [25]. In our case, timeslots can be viewed as resources, of capacity m, required by events. The contention C(j) of a time slot j is C(j) =
11 Table 3. A comparison of the various CP models we studied. Model 2 Model 3 (conflicts) Model 3 (all) Model 3 (room) Inst. iter time nodes iter time nodes iter time nodes iter time nodes Table 4. Comparison of search heuristics for the CP models. Impact Contention Dom/Deg Inst. iter time nodes iter time nodes iter time nodes i j D(eventT ime 1/ D(eventT ime i) i). Intuitively, this quantity describes the demand for timeslot j. It clearly induces a value ordering, since less contended for time slots are less likely to lead to a failure. Next we can compute a contention value for variables C(eventT ime i ), representing how constrained is a given variable and equal to 1 C(j). The event i that minimises C(eventT ime i) C(eventT ime i ) = j D(eventT ime i) and the timeslot j that minimises C(j) are explored first. In Table 4, we give the number iterations of Decompositon (Alg. 1) as well as the cumulative cpu time and number of nodes explored on solved instances. The results clearly show that contention dominates dom/deg and is itself dominated by impact. Notice that these two better heuristics also provide value orderings, whereas dom/deg does not. This is important on these benchmarks, since they have a relatively large number of solutions whilst being hard for a complete method. 5 Constraint Programming Models for Optimisation In this section we introduce three soft global constraints to reason about the costs and especially derive lower bounds. The main difficulty we encountered is that all three costs are defined in terms of students who are numerous and, thus, not represented explicitly in our CP model. In each case we tried to get around this issue by projecting the cost on events and/or timeslots. 5.1 Last Timeslot of each Day
12 This soft constraint counts the number of students attending an event in the last timeslot of the day ({9, 18, 27, 36, 45}). Let us introduce, for each event i a Boolean variable b i such that b i = 0 if the event i is in a timeslot other than the last ones, and b i = 1 if the event i is in one of the last timeslots. The cost can then be expressed as cost 1 = i (b i student(i) ). The Boolean variables can be added to Model 3 and channelled with eventt ime i or a simple dedicated global constraint can be implemented. We chose the latter option for efficiency reasons and to be able to augment it with stronger inference. Lower Bound. Consider the bipartite graph G = (V 1, V 2, E) described in Section 4.2 and captured by constraint (8) of Model 2. We recall that V 1 = {1,..., n} is the set of events and V 2 = { 1, 1,..., 45, m } is the set of all pairs timeslot, room. The existence of a maximum matching in this graph ensures a possible allocation of each event to a pair timeslot, room. W G extends G by adding a weight w ij to each edge of E defined as follows: { 0 iff j = a, b with a {9, 18, 27, 36, 45}; w ij = student(i) otherwise. Let us denote by W the value of the maximum weighted matching in W G. Observe that W represents the maximum number of students who can fit in the 40 nonlast timeslots and the rest is therefore a lower bound on the minimum number of student going in the last timeslots: lb(cost 1 ) = i n student(i) W. Pruning. The pruning process is trivial here and removes values {9, 18, 27, 36, 45} from the domain of an event eventt ime i if lb(cost 1 ) + student(i) > ub and event i has not been included in lb(cost 1 ). This can be done in O(n) time. Computational Complexity. The maximum weighted matching corresponds to an assignment problem and can be solved in polynomial time (in O(n 3 ) with the Hungarian method [16]). As n can reach 400 in the data sets, an incremental algorithm for the maximum weighted matching is needed and this improved bound has not yet been included in our current implementation. Note that this bound is exact when relaxing only constraints 1 and 5 of the problem description. We have seen that the colouring subproblem can however be tighter than the matchings so that reasoning on the colouring might improve this bound. 5.2 Consecutive Events This soft constraint counts the number of students attending more than two events in a row on a given day. The main difficulty with this cost is the potentially large number of parameters having an impact on the cost. We present the lower bound developed for this cost and how it is maintained incrementally at a relatively low computational cost. Lower Bound. We first consider only events committed to a timeslot, i.e., instantiated variables. The cost of consecutive allocation of every possible triplet of events is precomputed initially and stored in a large static table: staticcost(i 1, i 2, i 3 ) = student(i 1 ) student(i 2 ) student(i 3 ). The lower bound, lb(cost 2 ), is then, firstly,
13 made of the sum of these costs implied by instantiated events i 1, i 2, i 3 to consecutive timeslots. This part of the bound is referred to as lb g (cost 2 ): staticcost(i 1, i 2, i 3 ) if eventt ime i1/i 2/i 3 are groundcost(i 1, i 2, i 3 ) = assigned and consecutive; 0 otherwise. lb g (cost 2 ) = i 1<i 2<i 3 groundcost(i 1, i 2, i 3 ). Then, for each unassigned event, a lower bound on the cost involved by its insertion in the current timetable is maintained. For a timeslot j, let pairs(j) be the set of pairs of events assigned respectively to j 2 and j 1, or j 1 and j + 1, or j + 1 and j + 2. The cost of assigning event i to timeslot j is the sum of all triplets formed by i and any existing pair p in pairs(j). Then we define pendingcost(i, j) as: p pairs(j) staticcost(p {i}). The lower bound lb(i) associated with allocating event i to one of its possible timeslots is equal to the minimum pending cost over all values lb(i) = min j D(eventT imei)pendingcost(i, j). We use the following lower bound during search: lb(cost 2 ) = lb g (cost 2 ) + lb(i). D(eventT ime i) >1 Pruning. We prune timeslot j for event i iff lb(cost 2 ) + pendingcost(i, j) lb(i) is greater than the current upper bound of the variable associated to this cost. Computational Complexity. The base lower bound lb g (cost 2 ) is maintained incrementally during search. It is updated only when a variable eventt ime becomes assigned to some timeslot. In this case we increase the cost by the the value of staticcost of the newly formed triplets of events. There are at most 35m 3 triplets in total, the amortised computational cost of maintaining this lower bound along one branch of the search tree is thus O(m 3 ). The precomputation of the staticcost is here a key for efficiency. Computing pendingcost(i, j) can be done in O(m 2 ) time since there are three sets of at most m 2 pairs to consider for each timeslot of each event. Since there are at most 45 possible timeslots for a given event, one can compute lb(i) for all events in O(nm 2 ). In practice, we update the values of lb(i) only when event i loses some values, or when another variable get assigned to some timeslot j and D(eventT ime i ) {j 2, j 1, j + 1, j + 2}. The pruning can be done in the same time complexity since we only need to go through at most 45n values. Alternative Lower Bound. For a given student s, let us introduce the following Boolean variables: for each timeslot j, b s j = 1 if the student has an event assigned to the timeslot j, b s j = 0 if he is free at that time. These variables can be easily channelled with the eventt ime i variables. For a given assignment of b s 1,..., b s 45, c s is the corresponding cost (i.e. the sum of the number of triplets by day), and for a given c 0, opt s (c) = max{ j 45 bs j cs c}. In other words, opt s (c) is the biggest number of b s j variables we can set to 1 without exceeding the cost c.
14 A lower bound lb (s) of the cost for the student s can thus be defined as follows: lb (s) = min{c 0 events(s) opt s (c)} i.e. the minimal cost at which we can place all the events of the student s. The alternative lower bound for the cost is therefore lb (cost 2 ) = s S lb (s). Proposition 1. Let c be a value 0, in a given state of the b s i opt s (c) is polynomial. variables, determining Proof. Consider each timeslot of one day from the first to the last and the following greedy procedure. If an event can be added in the current timeslot without creating more triplets than c, add it, else, let this timeslot empty. Repeat this for each day. The number of events finally added is optimal. Indeed, suppose we are at the timeslot k in some day, and let a(k 1) = j k 1 bs j, a be the total number of bs j set to 1 when b s k = 1 and a when b s k = 0. Let us prove a. a If b s k = 1 increases the number of triplets over c, clearly we have no other choice than setting it to 0. Suppose it is not the case. If we set b s k to 1, then in the worst case, that is if b s k 1 = 1, we must let k + 1 empty and optimally fill the remaining timeslots with a r 1 s. Thus a a(k 1) a r. If we set b s k to 0, then the remaining slots can be optimally filled with a r 1 s, and so a = a(k 1) + a r. We have a r = a r + 1 or a r = a r, so a a(k 1) a r a(k 1) + a r a. So, supposing (by induction) that a(k 1) was optimal, we optimise the total number of 1 s by setting b s k to 1 whenever that is possible. This is of course not an exact lower bound, as we do not take into account that one event can only go to a single timeslot, as nothing prevents us from putting the same event into two different timeslots for two different students in order to optimise the cost of each of them. We also do not take into account the domains of the events. Proposition 2. lb(cost 2 ) and lb (cost 2 ) are incomparable. Proof. lb(cost 2 ) is not better than lb (cost 2 ): Consider three events 1, 2, 3 such that students(1) students(2) students(3) = {s}. Suppose that we have cost 2 0 and eventt ime 1, eventt ime 2 {1, 2}, eventt ime 3 {3}. We have lb g (cost 2 ) = 0, lb(1) = 0, lb(2) = 0, hence the overall lower bound is lb(cost 2 ) = 0. However using the alternative cost, when considering the student shared by all three events, we have b s 1 {0, 1}, b s 2 {0, 1}, b s 3 = {1}. Supposing we only have a single day with three timeslots (we can easily repeat this basic pattern to fill the whole week for a more realistic situation), lb (s) = 1, hence lb (cost 2 ) > 0. lb (cost 2 ) is not better than lb(cost 2 ): Suppose now we have two students indexed 1 and 2, such that the first one is busy at timeslots 1,2 and 6, and the second at timeslots 1, 5 and 6. We have one more event taken by both students that can go in 3 or 4. Then lb (1) = lb (2) = lb (cost 2 ) = 0, by setting for each student respectively b 1 3 = 0, b 1 4 = 1 and b 2 3 = 1, b 2 4 = 0. However this does obviously not lead to a solution, which would have been detected by lb(cost 2 ). Indeed, lb(3) = lb(4) = lb(cost 2 ) = 1.
15 5.3 Single Events This soft constraint counts students attending a single course in any day of the week. The nonmonotonic nature of this cost makes it difficult to reason about. Indeed, scheduling an event in a given day simultaneously increases the cost for students attending only this event, and decreases it for student attending another, until then unique, event. In fact, we show that even when we relax all other factors to an extremal case, computing an exact lower bound for this constraint is NPhard. Theorem 1. Finding the exact lower bound for Cost 3 is NPcomplete, even if all other constraints are relaxed. Proof. We consider the problem of finding a lower bound for cost 3 for a given day, with as few external constraints as possible. We only assume that a set of events may already have been assigned to this day, and that we have a finite set of extra events to choose from. We analyse the corresponding decision problem, SINGLEEVENT: Data: An integer k, a set R of events already assigned to a given day, and another set P that can possibly be assigned to this day. Question: Is there a set R S P of events such that no more than k students have a single event in that day. We reduce SETCOVER to SINGLEEVENT. A SETCOVER instance is composed of a set U = {u 1,..., u n }, a set S = {S 1,..., S M } 2 U of subsets of U, an integer k M. The problem consists of deciding whether there exists a set C S such that Si CS i = U and C k. We build R with one event E, that contains k + 1 students e 1 i,..., ek+1 i per element u i of U (the elementstudents). We build P with an event E j for each subset S j S. Each event E j contains the elementstudents of each element in S j, i.e. the elementstudents e 1 i,..., ek+1 i for each u i S j, plus one unique student s j (the setstudent). Each subset R S P of cost k, i.e. such that no more than k students attend a single event in the day, corresponds to a set cover of U of size k and viceversa. A set cover C of size k corresponds to a set S = {E j S j C} of cost k. Conversely, a set S of cost k corresponds to a set cover of size k. Clearly, S corresponds to a set cover: if any element of U is not covered, then the cost of S is at least k + 1 (each uncovered element corresponds to k + 1 elementstudents attending only E). Now, as all the elementstudents attend at least two events, the cost can only result from the setstudents, which is simply the number of events (other than E) in S. Observe that solving a sequence of SINGLEEVENT instances with decreasing values of k gives us a lower bound on this cost when all other constraints are relaxed, and without even imposing each variable to take at least one value. For instance if event i is in P, we can choose not to schedule it at all, whereas in effect, it will necessarily be assigned to some day of the week. This is, therefore, a much easier problem than finding the exact lower bound of Cost 3. However, even this relaxed problem is NPcomplete. Taking this fact into consideration, we only maintain this cost correctly in the computationally cheapest possible way. We consider each pair day, student. As long as at least two events attended by this student can potentially happen this day, we do nothing.
16 Otherwise, there are two choices, either this student has no course at all in this day, or has exactly one. In the latter case we increase the cost by one. This can be efficiently done with a system akin to the watched literals used in SAT unit propagation [18]. For every student and every day, we randomly pick two events to watch for this pair. When an event cannot be assigned to some day anymore, we update its list of students watched for that day, finding a new available watcher. Notice that this is very cheap to do. For instance if this event was not watching any student for that day, it does not cost anything at all. When we do not find any replacement, we know that the given student is either attending no event in that day, or only a single one. We update the cost accordingly. 6 Large Neighbourhood Search One weakness of the local search approach is the lack of flexibility when moving in the space of feasible solutions. The search space accessible from a given feasible solution might be very limited by the hard constraints and even disconnected. In such a case, the search can only reach the best solution connected with the initial one. One solution would be to relax feasibility during search without any guarantees to find it again or to restart from different feasible solutions. Due to the difficulty of finding feasible solutions, we discarded these two approaches. Another alternative is to design more complex moves that affect larger parts of the current assignment. Move Hu is one example of a complex move that remains polynomial. A more general kind of move can be performed using a complete solver. This is the central idea of Large Neighbourhood Search (LNS) [27]. LNS is a local search paradigm where the neighbourhood is defined by fixing a part of an existing solution. The rest of the variables are said to be released and all possible extensions of the fixed part define the neighbourhood which is usually much larger than the one obtained from classical and elementary moves. Algorithm 2 presents the simple LNS scheme. An efficient systematic algorithm is needed to explore this large neighbourhood and the CP Model 3 presented earlier will be used for this. Algorithm 2: LNS Scheme 1 find a feasible solution; 2 while optimal solution not found or time limit not reached do 3 choose a set of events to release; 4 freeze the remaining events to their current position; 5 if search for an improving solution then 6 update the upper bound; Nature and Size of the Neighbourhood. The selection of variables to release is a key element of a LNS scheme. We need to decide which events should be released (nature of the neighbourhood), and in what number (size of the neighbourhood). Previous work on LNS [7, 19, 20] outlines the importance of structured neighbourhoods dedicated to the problem. We have investigated neighbourhoods that release events per timeslots (all events contained in a given set of timeslots). It is critical to choose a neighbourhood
17 that releases related variables, i.e. variables that are likely to be able to change and exchange their values. It is indeed very important that the neighbourhood contain more feasible solutions than the one we already had before releasing the variables. A promising neighbourhood is also likely to contain feasible solutions of better cost. We, therefore, investigated a neighbourhood that releases kc conflicting and kr random timeslots (all events in the corresponding timeslots are released). The size of the neighbourhood is difficult to set as the tradeoff between searching more versus searching more often is difficult to achieve. We choose to start from small sizes (kc = 2 and kr = 2) and to increase it when the search stagnates; in practice, after 100 non improving iterations, the minimum of kc and kr is increased by 1. The reason is that the accurate size seems to vary a lot regarding instances. Much bigger sizes are typically needed for instances 1, 2, 9 and 10 where feasibility is tight. Sideways moves are again very important for diversification and are always accepted. LNS as an intensification mechanism for the SA. The LNS approach relies only on the CP solver as shown on Algorithm 2. Another idea is to use the LNS move at the low temperatures of the SA to help the very important and final phase of optimisation performed at the end of the cooling. In this mode, we do not accept sideways moves to speed up the solving process and look for improving solutions only. Diversification is ensured by other moves of the SA that continuously change the current assignment. The CP move is included in the neighbourhood of the SA at each iteration with a probability that increases while the temperature decreases: p include lns (τ) = τ. 7 Experimental Results Comparison of our Approaches. We summarise here the results of our study with four approaches: CP: The Constraint Programming approach described in Sections 4 and 5 based on Model 3 (room) using Impactbased search. LNS: The Large Neighbourhood Search approach relying on Model 3 (room) (the local search of Section 3.1 is used to provide an initial feasible solution). SA: The local search approach described in Section 3 which is based on Simulated Annealing for the optimisation stage. SA LNS: the SA approach augmented with LNS as an intensification mechanism at the end of the cooling (still using Model 3 (room)). Table 5 reports the cost found by each technique on the 16 instances. LNS, SA and SA LNS were run on 20 different seeds and the average, min and max cost found over the 20 runs are reported. The CP approach is entirely deterministic and a single run is therefore shown. The last three columns show the percentage of improvements given by SA LNS over SA alone. Two computers were used, CP was run on a Mac Book 4 within a time limit of 420s and the others were run on an imac 5 within a time 4 Mac OS X , 2 GHz Intel Core 2 Duo, 2 GB 667 Mhz DDR2. 5 Mac OS X , 2.33 GHz Intel Core 2 Duo, 3 GB 667MHz DDR2 SDRAM.
18 limit of 372s 6. Firstly, the LNS scheme outperforms CP alone (even by only looking at instances where CP does find a feasible solution) while being a very simple modification of CP. Secondly, LNS is itself outperformed by the SA. We observed here that it remains stuck in local minima despite the large size of the neighbourhood. [19] outlines the same problem and suggests that the LNS scheme still needs other and more powerful diversification mechanisms. Finally, SA LNS improves LNS but is not very convincing. The CP moves do not seem to bring much more flexibility to the SA to escape local minima in general. It allows, however, to find three new optimal solutions (instances 7, 12 and 16) and improves significantly the resolution of two instances (7 and 16). Note that all the minimum costs are improved showing that LNS does play a role in the final intensification stage even if this does not give a major improvement. Table 5. Overall results on 20 runs reporting the average, min and max cost. Inst CP LNS SA SA LNS Improvements avg min max avg min max avg min max %avg %min %max Comparisons with other Algorithms in the Competition. Five algorithms 7 were choosen for the final phase and evaluated on 24 instances (the 16 already mentioned and 8 unknown competition ones). Since all solvers were randomised, 10 runs per instance were performed giving 50 runs per instance. Each run was ranked among the 50 for each instance and the average rank across all runs and all instances was used to give a rank to each algorithm. Table 6 shows the ranking of each algorithm, with the number of times they have found the best solution among all runs for a given instance, and the number of times they have failed to find a feasible solution. Our local search with a score of 13.9 therefore did significantly better than the algorithm of Atsuta et al. that came second with Our approach appeared generally more robust for both finding feasible and good solutions (Chirandini et al. being the most robust on feasibility only). It also obtained the best results on many instances. It is however very interesting to notice that it was outperformed on instances we considered here as very hard on feasibility for our algorithm, e.g. instance 10. Nothegger et al. or 6 Both time limits were established by the benchmarking program used during the competition. 7 The other four algorithms were developped by: M. Atsuta, K. Nonobe, T. Ibaraki; M. Chiarandini, C. Fawcett, H. H Hoos; C. Nothegger, A. Mayer, A. Chwatal, G. Raidi; T.Muller.
19 Table 6. Ranking of the five finalists from the tests ran by the organizers on the 24 instances. Atsuta et al. Cambazard et al. Chiarandini et al. Nothegger et al. Muller Average rank (out of 240 runs) Number of best solutions (out of 24 instances) Number of failures on feasibility (out of 240 runs) Rank in the competition Atsuki et al. could not only systematically find a feasible solution on instance 10 but also the optimal one. On the other hand, these algorithms fail to find feasible solutions on instances where our approach succeeds easily. On instance 9 they only find a feasible solution 30% of the time but when they do, it is the optimal one. This lead us to conjecture that they have been using the soft cost to guide the search of feasible solutions. On very tight instances, this strategy pays off as there are maybe few feasible solutions and it is known that an optimal solution of cost 0 always exists. On other instances it either misleads the search or just slows down the process, thus degrading the results. This shows that there is a significant room for improvement in our results. 8 Conclusion We have presented a comprehensive study of a university timetabling problem, comparing a variety of local search and constraint programming approaches. We designed a constraint programming approach that proceeds by decomposing the listcolouring and the matching subproblems and outperforms more classical CP models. Lower bounds were introduced to tackle soft constraints, leading to the first complete algorithm for this problem. While our local search technique benefits from the experience of the 2003 competition, we have presented several improvements to deal with hard constraints; the results show more maturity than the CP technique. However, an LNS scheme integrating both our CP and LS approaches obtained the best results. The structure of the listcolouring graph made of large and overlapping cliques was shown to be important for both CP and LS techniques. Improving the propagation we can achieve from a collection of ALLDIFF constraints is very important in this context. Arcconsistency on two overlapping ALLDIFF is already known to be NPComplete [3] but a number of pragmatic filtering rules could be designed. This is an important topic for future work. References 1. S. Abdullah, E. K. Burke, and B. McCollum. Using a randomised iterative improvement algorithm with composite neighbourhood structures for course timetabling. In MIC 05: The 6th MetaHeuristic International Conference, A. Aggoun and N. Beldiceanu. Extending chip in order to solve complex scheduling and placement problems. Mathematical Computing and Modelling, 17(7):57 73, N. Beldiceanu, M. Carlsson, S. Demassey, and T. Petit. Global constraint catalogue: Past, present and future. Constraints, 12(1):21 62, H. Cambazard, P.E. Hladik, A.M. Déplanche, N. Jussien, and Y. Trinquet. Decomposition and learning for a real time task allocation problem. In Proc. of CP, pages , M. W. Carter and G. Laporte. Recent developments in practical course timetabling. In PATAT, pages 3 19, 1997.
20 6. M. Chiarandini, M. Birattari, K. Socha, and O. RossiDoria. An effective hybrid algorithm for university course timetabling. J. Scheduling, 9(5): , E. Danna and L. Perron. Structured vs. unstructured large neighborhood search: A case study on jobshop scheduling problems with earliness and tardiness costs. In CP, pages , J.L. de Siqueira and J.F. Puget. Explanationbased generalisation of failures. In European Conference on Artificial Intelligence (ECAI 88), pages , D. de Werra. An introduction to timetabling. European Journal of Operational Research, 19(2): , February L. Di Gaspero and A. Schaerf. Neighborhood portfolio approach for local search applied to timetabling problems. Journal of Mathematical Modeling and Algorithms, 5(1):65 89, P. Galinier and A. Hertz. A survey of local search methods for graph coloring. Comput. Oper. Res., 33(9): , J.N. Hooker and G. Ottosson. Logicbased benders decomposition. Mathematical Programming, 96:33 60, V. Jain and I. E. Grossmann. Algorithms for hybrid milp/cp models for a class of optimization problems. INFORMS Journal on Computing, 13: , S. Kirkpatrick, C. D. Gelatt, and M. P. Vecchi. Optimization by simulated annealing. Science, Number 4598, 13 May 1983, 220, 4598: , P. Kostuch. The university course timetabling problem with a threephase approach. In PATAT, pages , H. W. Kuhn. The hungarian method for the assignment problem. Naval Research Logistics Quarterly, 2(1):83 98, R. Lewis, B. Paechter, and B. McCollum. Post enrolment based course timetabling: A description of the problem model used for track two of the second international timetabling competition. Technical report, Cardiff University, M. W. Moskewicz, C. F. Madigan, Y. Zhao, L. Zhang, and S. Malik. Chaff: Engineering an Efficient SAT Solver. In Proceedings of the 38th Design Automation Conference (DAC 01), pages , L. Perron and P. Shaw. Combining forces to solve the car sequencing problem. In CPAIOR, pages , L. Perron, P. Shaw, and V. Furnon. Propagation guided large neighborhood search. In CP, pages , P. Refalo. Impactbased search strategies for constraint programming. In CP, pages , J.C. Régin. A filtering algorithm for constraints of difference in CSPs. In Proceedings of the 12th National Conference on Artificial Intelligence (AAAI94), pages , J.C. Régin. Generalized arc consistency for global cardinality constraint. In National Conference on Artificial Intelligence (AAAI 96), pages , O. RossiDoria, M. Sampels, M. Birattari, M. Chiarandini, M. Dorigo, L. M. Gambardella, J. D. Knowles, M. Manfrin, M. Mastrolilli, B. Paechter, L. Paquete, and T. Stützle. A comparison of the performance of different metaheuristics on the timetabling problem. In PATAT, pages , N. Sadeh and M.S. Fox. Variable and Value Ordering Heuristics for the JobShop Scheduling Constraint Satisfaction Problem. Artificial Intelligence, 86(1):1 41, September A. Schaerf. A survey of automated timetabling. Artificial Intelligence Review, 13(2):87 127, P. Shaw. Using constraint programming and local search methods to solve vehicle routing problems. In CP, pages , 1998.
Local Search and Constraint Programming for the Post Enrolmentbased Course Timetabling Problem
Local Search and Constraint Programming for the Post Enrolmentbased Course Timetabling Problem Hadrien Cambazard, Emmanuel Hebrard, Barry O Sullivan and Alexandre Papadopoulos Cork Constraint Computation
More informationA Constraint Programming based Column Generation Approach to Nurse Rostering Problems
Abstract A Constraint Programming based Column Generation Approach to Nurse Rostering Problems Fang He and Rong Qu The Automated Scheduling, Optimisation and Planning (ASAP) Group School of Computer Science,
More informationIntegrating Benders decomposition within Constraint Programming
Integrating Benders decomposition within Constraint Programming Hadrien Cambazard, Narendra Jussien email: {hcambaza,jussien}@emn.fr École des Mines de Nantes, LINA CNRS FRE 2729 4 rue Alfred Kastler BP
More informationBranchandPrice Approach to the Vehicle Routing Problem with Time Windows
TECHNISCHE UNIVERSITEIT EINDHOVEN BranchandPrice Approach to the Vehicle Routing Problem with Time Windows Lloyd A. Fasting May 2014 Supervisors: dr. M. Firat dr.ir. M.A.A. Boon J. van Twist MSc. Contents
More informationDiscuss the size of the instance for the minimum spanning tree problem.
3.1 Algorithm complexity The algorithms A, B are given. The former has complexity O(n 2 ), the latter O(2 n ), where n is the size of the instance. Let n A 0 be the size of the largest instance that can
More informationScheduling Shop Scheduling. Tim Nieberg
Scheduling Shop Scheduling Tim Nieberg Shop models: General Introduction Remark: Consider non preemptive problems with regular objectives Notation Shop Problems: m machines, n jobs 1,..., n operations
More informationScheduling Home Health Care with Separating Benders Cuts in Decision Diagrams
Scheduling Home Health Care with Separating Benders Cuts in Decision Diagrams André Ciré University of Toronto John Hooker Carnegie Mellon University INFORMS 2014 Home Health Care Home health care delivery
More informationSingleLink Failure Detection in AllOptical Networks Using Monitoring Cycles and Paths
SingleLink Failure Detection in AllOptical Networks Using Monitoring Cycles and Paths Satyajeet S. Ahuja, Srinivasan Ramasubramanian, and Marwan Krunz Department of ECE, University of Arizona, Tucson,
More informationThe Second International Timetabling Competition (ITC2007): Curriculumbased Course Timetabling (Track 3)
The Second International Timetabling Competition (ITC2007): Curriculumbased Course Timetabling (Track 3) preliminary presentation Luca Di Gaspero and Andrea Schaerf DIEGM, University of Udine via delle
More informationExponential time algorithms for graph coloring
Exponential time algorithms for graph coloring Uriel Feige Lecture notes, March 14, 2011 1 Introduction Let [n] denote the set {1,..., k}. A klabeling of vertices of a graph G(V, E) is a function V [k].
More informationThe Problem of Scheduling Technicians and Interventions in a Telecommunications Company
The Problem of Scheduling Technicians and Interventions in a Telecommunications Company Sérgio Garcia Panzo Dongala November 2008 Abstract In 2007 the challenge organized by the French Society of Operational
More informationY. Xiang, Constraint Satisfaction Problems
Constraint Satisfaction Problems Objectives Constraint satisfaction problems Backtracking Iterative improvement Constraint propagation Reference Russell & Norvig: Chapter 5. 1 Constraints Constraints are
More informationSchool Timetabling in Theory and Practice
School Timetabling in Theory and Practice Irving van Heuven van Staereling VU University, Amsterdam Faculty of Sciences December 24, 2012 Preface At almost every secondary school and university, some
More informationThe Classes P and NP
The Classes P and NP We now shift gears slightly and restrict our attention to the examination of two families of problems which are very important to computer scientists. These families constitute the
More informationSmart Graphics: Methoden 3 Suche, Constraints
Smart Graphics: Methoden 3 Suche, Constraints Vorlesung Smart Graphics LMU München Medieninformatik Butz/Boring Smart Graphics SS2007 Methoden: Suche 2 Folie 1 Themen heute Suchverfahren Hillclimbing Simulated
More informationApproximation Algorithms
Approximation Algorithms or: How I Learned to Stop Worrying and Deal with NPCompleteness Ong Jit Sheng, Jonathan (A0073924B) March, 2012 Overview Key Results (I) General techniques: Greedy algorithms
More informationA ConstraintBased Method for Project Scheduling with Time Windows
A ConstraintBased Method for Project Scheduling with Time Windows Amedeo Cesta 1 and Angelo Oddi 1 and Stephen F. Smith 2 1 ISTCCNR, National Research Council of Italy Viale Marx 15, I00137 Rome, Italy,
More informationComplexity Theory. IE 661: Scheduling Theory Fall 2003 Satyaki Ghosh Dastidar
Complexity Theory IE 661: Scheduling Theory Fall 2003 Satyaki Ghosh Dastidar Outline Goals Computation of Problems Concepts and Definitions Complexity Classes and Problems Polynomial Time Reductions Examples
More informationWhy? A central concept in Computer Science. Algorithms are ubiquitous.
Analysis of Algorithms: A Brief Introduction Why? A central concept in Computer Science. Algorithms are ubiquitous. Using the Internet (sending email, transferring files, use of search engines, online
More informationThe Trip Scheduling Problem
The Trip Scheduling Problem Claudia Archetti Department of Quantitative Methods, University of Brescia Contrada Santa Chiara 50, 25122 Brescia, Italy Martin Savelsbergh School of Industrial and Systems
More informationSatisfiability Checking
Satisfiability Checking SATSolving Prof. Dr. Erika Ábrahám Theory of Hybrid Systems Informatik 2 WS 10/11 Prof. Dr. Erika Ábrahám  Satisfiability Checking 1 / 40 A basic SAT algorithm Assume the CNF
More informationClustering and scheduling maintenance tasks over time
Clustering and scheduling maintenance tasks over time Per Kreuger 20080429 SICS Technical Report T2008:09 Abstract We report results on a maintenance scheduling problem. The problem consists of allocating
More informationOptimizing Description Logic Subsumption
Topics in Knowledge Representation and Reasoning Optimizing Description Logic Subsumption Maryam FazelZarandi Company Department of Computer Science University of Toronto Outline Introduction Optimization
More informationCSC 373: Algorithm Design and Analysis Lecture 16
CSC 373: Algorithm Design and Analysis Lecture 16 Allan Borodin February 25, 2013 Some materials are from Stephen Cook s IIT talk and Keven Wayne s slides. 1 / 17 Announcements and Outline Announcements
More information5.1 Bipartite Matching
CS787: Advanced Algorithms Lecture 5: Applications of Network Flow In the last lecture, we looked at the problem of finding the maximum flow in a graph, and how it can be efficiently solved using the FordFulkerson
More informationINTEGER PROGRAMMING. Integer Programming. Prototype example. BIP model. BIP models
Integer Programming INTEGER PROGRAMMING In many problems the decision variables must have integer values. Example: assign people, machines, and vehicles to activities in integer quantities. If this is
More informationCompact Representations and Approximations for Compuation in Games
Compact Representations and Approximations for Compuation in Games Kevin Swersky April 23, 2008 Abstract Compact representations have recently been developed as a way of both encoding the strategic interactions
More informationnpsolver A SAT Based Solver for Optimization Problems
npsolver A SAT Based Solver for Optimization Problems Norbert Manthey and Peter Steinke Knowledge Representation and Reasoning Group Technische Universität Dresden, 01062 Dresden, Germany peter@janeway.inf.tudresden.de
More informationTowards Optimal Firewall Rule Ordering Utilizing Directed Acyclical Graphs
Towards Optimal Firewall Rule Ordering Utilizing Directed Acyclical Graphs Ashish Tapdiya and Errin W. Fulp Department of Computer Science Wake Forest University Winston Salem, NC, USA nsg.cs.wfu.edu Email:
More informationDynamic programming. Doctoral course Optimization on graphs  Lecture 4.1. Giovanni Righini. January 17 th, 2013
Dynamic programming Doctoral course Optimization on graphs  Lecture.1 Giovanni Righini January 1 th, 201 Implicit enumeration Combinatorial optimization problems are in general NPhard and we usually
More informationOffline sorting buffers on Line
Offline sorting buffers on Line Rohit Khandekar 1 and Vinayaka Pandit 2 1 University of Waterloo, ON, Canada. email: rkhandekar@gmail.com 2 IBM India Research Lab, New Delhi. email: pvinayak@in.ibm.com
More informationTHE PROBLEM WORMS (1) WORMS (2) THE PROBLEM OF WORM PROPAGATION/PREVENTION THE MINIMUM VERTEX COVER PROBLEM
1 THE PROBLEM OF WORM PROPAGATION/PREVENTION I.E. THE MINIMUM VERTEX COVER PROBLEM Prof. Tiziana Calamoneri Network Algorithms A.y. 2014/15 2 THE PROBLEM WORMS (1)! A computer worm is a standalone malware
More informationFairness in Routing and Load Balancing
Fairness in Routing and Load Balancing Jon Kleinberg Yuval Rabani Éva Tardos Abstract We consider the issue of network routing subject to explicit fairness conditions. The optimization of fairness criteria
More informationeach college c i C has a capacity q i  the maximum number of students it will admit
n colleges in a set C, m applicants in a set A, where m is much larger than n. each college c i C has a capacity q i  the maximum number of students it will admit each college c i has a strict order i
More informationGuessing Game: NPComplete?
Guessing Game: NPComplete? 1. LONGESTPATH: Given a graph G = (V, E), does there exists a simple path of length at least k edges? YES 2. SHORTESTPATH: Given a graph G = (V, E), does there exists a simple
More information! Solve problem to optimality. ! Solve problem in polytime. ! Solve arbitrary instances of the problem. #approximation algorithm.
Approximation Algorithms 11 Approximation Algorithms Q Suppose I need to solve an NPhard problem What should I do? A Theory says you're unlikely to find a polytime algorithm Must sacrifice one of three
More informationAdaptive Linear Programming Decoding
Adaptive Linear Programming Decoding Mohammad H. Taghavi and Paul H. Siegel ECE Department, University of California, San Diego Email: (mtaghavi, psiegel)@ucsd.edu ISIT 2006, Seattle, USA, July 9 14, 2006
More informationInformation Theory and Coding Prof. S. N. Merchant Department of Electrical Engineering Indian Institute of Technology, Bombay
Information Theory and Coding Prof. S. N. Merchant Department of Electrical Engineering Indian Institute of Technology, Bombay Lecture  17 ShannonFanoElias Coding and Introduction to Arithmetic Coding
More informationClassification  Examples
Lecture 2 Scheduling 1 Classification  Examples 1 r j C max given: n jobs with processing times p 1,...,p n and release dates r 1,...,r n jobs have to be scheduled without preemption on one machine taking
More information! Solve problem to optimality. ! Solve problem in polytime. ! Solve arbitrary instances of the problem. !approximation algorithm.
Approximation Algorithms Chapter Approximation Algorithms Q Suppose I need to solve an NPhard problem What should I do? A Theory says you're unlikely to find a polytime algorithm Must sacrifice one of
More informationOperations and Supply Chain Management Prof. G. Srinivasan Department of Management Studies Indian Institute of Technology, Madras
Operations and Supply Chain Management Prof. G. Srinivasan Department of Management Studies Indian Institute of Technology, Madras Lecture  36 Location Problems In this lecture, we continue the discussion
More informationMapReduce and Distributed Data Analysis. Sergei Vassilvitskii Google Research
MapReduce and Distributed Data Analysis Google Research 1 Dealing With Massive Data 2 2 Dealing With Massive Data Polynomial Memory Sublinear RAM Sketches External Memory Property Testing 3 3 Dealing With
More informationCSC2420 Fall 2012: Algorithm Design, Analysis and Theory
CSC2420 Fall 2012: Algorithm Design, Analysis and Theory Allan Borodin November 15, 2012; Lecture 10 1 / 27 Randomized online bipartite matching and the adwords problem. We briefly return to online algorithms
More informationUniversity of Potsdam Faculty of Computer Science. Clause Learning in SAT Seminar Automatic Problem Solving WS 2005/06
University of Potsdam Faculty of Computer Science Clause Learning in SAT Seminar Automatic Problem Solving WS 2005/06 Authors: Richard Tichy, Thomas Glase Date: 25th April 2006 Contents 1 Introduction
More informationGlobally Optimal Crowdsourcing Quality Management
Globally Optimal Crowdsourcing Quality Management Akash Das Sarma Stanford University akashds@stanford.edu Aditya G. Parameswaran University of Illinois (UIUC) adityagp@illinois.edu Jennifer Widom Stanford
More informationMethods for Firewall Policy Detection and Prevention
Methods for Firewall Policy Detection and Prevention Hemkumar D Asst Professor Dept. of Computer science and Engineering Sharda University, Greater Noida NCR Mohit Chugh B.tech (Information Technology)
More informationRandomization Approaches for Network Revenue Management with Customer Choice Behavior
Randomization Approaches for Network Revenue Management with Customer Choice Behavior Sumit Kunnumkal Indian School of Business, Gachibowli, Hyderabad, 500032, India sumit kunnumkal@isb.edu March 9, 2011
More informationA hybrid approach for solving realworld nurse rostering problems
Presentation at CP 2011: A hybrid approach for solving realworld nurse rostering problems Martin Stølevik (martin.stolevik@sintef.no) Tomas Eric Nordlander (tomas.nordlander@sintef.no) Atle Riise (atle.riise@sintef.no)
More informationPolicy Analysis for Administrative Role Based Access Control without Separate Administration
Policy nalysis for dministrative Role Based ccess Control without Separate dministration Ping Yang Department of Computer Science, State University of New York at Binghamton, US Mikhail I. Gofman Department
More informationIE 680 Special Topics in Production Systems: Networks, Routing and Logistics*
IE 680 Special Topics in Production Systems: Networks, Routing and Logistics* Rakesh Nagi Department of Industrial Engineering University at Buffalo (SUNY) *Lecture notes from Network Flows by Ahuja, Magnanti
More informationNan Kong, Andrew J. Schaefer. Department of Industrial Engineering, Univeristy of Pittsburgh, PA 15261, USA
A Factor 1 2 Approximation Algorithm for TwoStage Stochastic Matching Problems Nan Kong, Andrew J. Schaefer Department of Industrial Engineering, Univeristy of Pittsburgh, PA 15261, USA Abstract We introduce
More informationEnergy Efficient Monitoring in Sensor Networks
Energy Efficient Monitoring in Sensor Networks Amol Deshpande, Samir Khuller, Azarakhsh Malekian, Mohammed Toossi Computer Science Department, University of Maryland, A.V. Williams Building, College Park,
More informationHighperformance local search for planning maintenance of EDF nuclear park
Highperformance local search for planning maintenance of EDF nuclear park Frédéric Gardi Karim Nouioua Bouygues elab, Paris fgardi@bouygues.com Laboratoire d'informatique Fondamentale  CNRS UMR 6166,
More informationOn the TradeOff between Control Plane Load and Data Plane Efficiency in Software Defined Networks
1 Technion  Computer Science Department  Tehnical Report CS010  01 On the TradeOff between Control Plane Load and Data Plane Efficiency in Software Defined Networks Abstract Software Defined Networking
More informationAnt Colony Optimization and Constraint Programming
Ant Colony Optimization and Constraint Programming Christine Solnon Series Editor Narendra Jussien WILEY Table of Contents Foreword Acknowledgements xi xiii Chapter 1. Introduction 1 1.1. Overview of the
More informationHow To Solve The Stable Roommates Problem
THE ROOMMATES PROBLEM DISCUSSED NATHAN SCHULZ Abstract. The stable roommates problem as originally posed by Gale and Shapley [1] in 1962 involves a single set of even cardinality 2n, each member of which
More informationOPRE 6201 : 2. Simplex Method
OPRE 6201 : 2. Simplex Method 1 The Graphical Method: An Example Consider the following linear program: Max 4x 1 +3x 2 Subject to: 2x 1 +3x 2 6 (1) 3x 1 +2x 2 3 (2) 2x 2 5 (3) 2x 1 +x 2 4 (4) x 1, x 2
More informationIntroduction to Logic in Computer Science: Autumn 2006
Introduction to Logic in Computer Science: Autumn 2006 Ulle Endriss Institute for Logic, Language and Computation University of Amsterdam Ulle Endriss 1 Plan for Today Now that we have a basic understanding
More informationThis article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
IEEE/ACM TRANSACTIONS ON NETWORKING 1 A Greedy Link Scheduler for Wireless Networks With Gaussian MultipleAccess and Broadcast Channels Arun Sridharan, Student Member, IEEE, C Emre Koksal, Member, IEEE,
More informationHeuristics for Dynamically Adapting Constraint Propagation in Constraint Programming
Heuristics for Dynamically Adapting Constraint Propagation in Constraint Programming Kostas Stergiou AI Lab University of the Aegean Greece CPAIOR 09 Workshop on Bound reduction techniques for CP and MINLP
More informationStudent Project Allocation Using Integer Programming
IEEE TRANSACTIONS ON EDUCATION, VOL. 46, NO. 3, AUGUST 2003 359 Student Project Allocation Using Integer Programming A. A. Anwar and A. S. Bahaj, Member, IEEE Abstract The allocation of projects to students
More informationA Branch and Bound Algorithm for Solving the Binary Bilevel Linear Programming Problem
A Branch and Bound Algorithm for Solving the Binary Bilevel Linear Programming Problem John Karlof and Peter Hocking Mathematics and Statistics Department University of North Carolina Wilmington Wilmington,
More informationA Mathematical Programming Solution to the Mars Express Memory Dumping Problem
A Mathematical Programming Solution to the Mars Express Memory Dumping Problem Giovanni Righini and Emanuele Tresoldi Dipartimento di Tecnologie dell Informazione Università degli Studi di Milano Via Bramante
More informationThe Relative Worst Order Ratio for OnLine Algorithms
The Relative Worst Order Ratio for OnLine Algorithms Joan Boyar 1 and Lene M. Favrholdt 2 1 Department of Mathematics and Computer Science, University of Southern Denmark, Odense, Denmark, joan@imada.sdu.dk
More informationA General Purpose Local Search Algorithm for Binary Optimization
A General Purpose Local Search Algorithm for Binary Optimization Dimitris Bertsimas, Dan Iancu, Dmitriy Katz Operations Research Center, Massachusetts Institute of Technology, E40147, Cambridge, MA 02139,
More informationFAST LOCAL SEARCH FOR THE MAXIMUM INDEPENDENT SET PROBLEM
FAST LOCAL SEARCH FOR THE MAXIMUM INDEPENDENT SET PROBLEM DIOGO V. ANDRADE, MAURICIO G.C. RESENDE, AND RENATO F. WERNECK Abstract. Given a graph G = (V, E), the independent set problem is that of finding
More informationAdaptive Tolerance Algorithm for Distributed TopK Monitoring with Bandwidth Constraints
Adaptive Tolerance Algorithm for Distributed TopK Monitoring with Bandwidth Constraints Michael Bauer, Srinivasan Ravichandran University of WisconsinMadison Department of Computer Sciences {bauer, srini}@cs.wisc.edu
More informationFactoring & Primality
Factoring & Primality Lecturer: Dimitris Papadopoulos In this lecture we will discuss the problem of integer factorization and primality testing, two problems that have been the focus of a great amount
More informationAnalysis of Approximation Algorithms for kset Cover using FactorRevealing Linear Programs
Analysis of Approximation Algorithms for kset Cover using FactorRevealing Linear Programs Stavros Athanassopoulos, Ioannis Caragiannis, and Christos Kaklamanis Research Academic Computer Technology Institute
More informationAn optimisation framework for determination of capacity in railway networks
CASPT 2015 An optimisation framework for determination of capacity in railway networks Lars Wittrup Jensen Abstract Within the railway industry, high quality estimates on railway capacity is crucial information,
More informationA Brief Study of the Nurse Scheduling Problem (NSP)
A Brief Study of the Nurse Scheduling Problem (NSP) Lizzy Augustine, Morgan Faer, Andreas Kavountzis, Reema Patel Submitted Tuesday December 15, 2009 0. Introduction and Background Our interest in the
More informationP versus NP, and More
1 P versus NP, and More Great Ideas in Theoretical Computer Science Saarland University, Summer 2014 If you have tried to solve a crossword puzzle, you know that it is much harder to solve it than to verify
More informationMinesweeper as a Constraint Satisfaction Problem
Minesweeper as a Constraint Satisfaction Problem by Chris Studholme Introduction To Minesweeper Minesweeper is a simple one player computer game commonly found on machines with popular operating systems
More informationNetwork Algorithms for Homeland Security
Network Algorithms for Homeland Security Mark Goldberg and Malik MagdonIsmail Rensselaer Polytechnic Institute September 27, 2004. Collaborators J. Baumes, M. Krishmamoorthy, N. Preston, W. Wallace. Partially
More informationSoftware Project Management with GAs
Software Project Management with GAs Enrique Alba, J. Francisco Chicano University of Málaga, Grupo GISUM, Departamento de Lenguajes y Ciencias de la Computación, E.T.S Ingeniería Informática, Campus de
More informationCost Models for Vehicle Routing Problems. 8850 Stanford Boulevard, Suite 260 R. H. Smith School of Business
0769514359/02 $17.00 (c) 2002 IEEE 1 Cost Models for Vehicle Routing Problems John Sniezek Lawerence Bodin RouteSmart Technologies Decision and Information Technologies 8850 Stanford Boulevard, Suite
More informationNear Optimal Solutions
Near Optimal Solutions Many important optimization problems are lacking efficient solutions. NPComplete problems unlikely to have polynomial time solutions. Good heuristics important for such problems.
More informationOutline. NPcompleteness. When is a problem easy? When is a problem hard? Today. Euler Circuits
Outline NPcompleteness Examples of Easy vs. Hard problems Euler circuit vs. Hamiltonian circuit Shortest Path vs. Longest Path 2pairs sum vs. general Subset Sum Reducing one problem to another Clique
More information4.2 Description of the Event operation Network (EON)
Integrated support system for planning and scheduling... 2003/4/24 page 39 #65 Chapter 4 The EON model 4. Overview The present thesis is focused in the development of a generic scheduling framework applicable
More informationLoad Balancing. Load Balancing 1 / 24
Load Balancing Backtracking, branch & bound and alphabeta pruning: how to assign work to idle processes without much communication? Additionally for alphabeta pruning: implementing the youngbrotherswait
More informationTHE SCHEDULING OF MAINTENANCE SERVICE
THE SCHEDULING OF MAINTENANCE SERVICE Shoshana Anily Celia A. Glass Refael Hassin Abstract We study a discrete problem of scheduling activities of several types under the constraint that at most a single
More informationLoad Balancing and Switch Scheduling
EE384Y Project Final Report Load Balancing and Switch Scheduling Xiangheng Liu Department of Electrical Engineering Stanford University, Stanford CA 94305 Email: liuxh@systems.stanford.edu Abstract Load
More informationThe Minimum Consistent Subset Cover Problem and its Applications in Data Mining
The Minimum Consistent Subset Cover Problem and its Applications in Data Mining Byron J Gao 1,2, Martin Ester 1, JinYi Cai 2, Oliver Schulte 1, and Hui Xiong 3 1 School of Computing Science, Simon Fraser
More informationM. Sugumaran / (IJCSIT) International Journal of Computer Science and Information Technologies, Vol. 2 (3), 2011, 10011006
A Design of Centralized Meeting Scheduler with Distance Metrics M. Sugumaran Department of Computer Science and Engineering,Pondicherry Engineering College, Puducherry, India. Abstract Meeting scheduling
More informationTutorial: Operations Research in Constraint Programming
Tutorial: Operations Research in Constraint Programming John Hooker Carnegie Mellon University May 2009 Revised June 2009 May 2009 Slide 1 Motivation Benders decomposition allows us to apply CP and OR
More informationMATH10212 Linear Algebra. Systems of Linear Equations. Definition. An ndimensional vector is a row or a column of n numbers (or letters): a 1.
MATH10212 Linear Algebra Textbook: D. Poole, Linear Algebra: A Modern Introduction. Thompson, 2006. ISBN 0534405967. Systems of Linear Equations Definition. An ndimensional vector is a row or a column
More informationScheduling Algorithm with Optimization of Employee Satisfaction
Washington University in St. Louis Scheduling Algorithm with Optimization of Employee Satisfaction by Philip I. Thomas Senior Design Project http : //students.cec.wustl.edu/ pit1/ Advised By Associate
More informationMINIMAL BOOKS OF RATIONALES
MINIMAL BOOKS OF RATIONALES José Apesteguía Miguel A. Ballester D.T.2005/01 MINIMAL BOOKS OF RATIONALES JOSE APESTEGUIA AND MIGUEL A. BALLESTER Abstract. Kalai, Rubinstein, and Spiegler (2002) propose
More informationChapter 11. 11.1 Load Balancing. Approximation Algorithms. Load Balancing. Load Balancing on 2 Machines. Load Balancing: Greedy Scheduling
Approximation Algorithms Chapter Approximation Algorithms Q. Suppose I need to solve an NPhard problem. What should I do? A. Theory says you're unlikely to find a polytime algorithm. Must sacrifice one
More informationResearch Statement Immanuel Trummer www.itrummer.org
Research Statement Immanuel Trummer www.itrummer.org We are collecting data at unprecedented rates. This data contains valuable insights, but we need complex analytics to extract them. My research focuses
More information9th MaxPlanck Advanced Course on the Foundations of Computer Science (ADFOCS) PrimalDual Algorithms for Online Optimization: Lecture 1
9th MaxPlanck Advanced Course on the Foundations of Computer Science (ADFOCS) PrimalDual Algorithms for Online Optimization: Lecture 1 Seffi Naor Computer Science Dept. Technion Haifa, Israel Introduction
More informationDiscrete Optimization
Discrete Optimization [Chen, Batson, Dang: Applied integer Programming] Chapter 3 and 4.14.3 by Johan Högdahl and Victoria Svedberg Seminar 2, 20150331 Todays presentation Chapter 3 Transforms using
More informationSYSM 6304: Risk and Decision Analysis Lecture 5: Methods of Risk Analysis
SYSM 6304: Risk and Decision Analysis Lecture 5: Methods of Risk Analysis M. Vidyasagar Cecil & Ida Green Chair The University of Texas at Dallas Email: M.Vidyasagar@utdallas.edu October 17, 2015 Outline
More informationIntroduction & Overview
ID2204: Constraint Programming Introduction & Overview Lecture 01, Christian Schulte cschulte@kth.se Software and Computer Systems School of Information and Communication Technology KTH Royal Institute
More information24. The Branch and Bound Method
24. The Branch and Bound Method It has serious practical consequences if it is known that a combinatorial problem is NPcomplete. Then one can conclude according to the present state of science that no
More informationMultilayer MPLS Network Design: the Impact of Statistical Multiplexing
Multilayer MPLS Network Design: the Impact of Statistical Multiplexing Pietro Belotti, Antonio Capone, Giuliana Carello, Federico Malucelli Tepper School of Business, Carnegie Mellon University, Pittsburgh
More informationMax Flow, Min Cut, and Matchings (Solution)
Max Flow, Min Cut, and Matchings (Solution) 1. The figure below shows a flow network on which an st flow is shown. The capacity of each edge appears as a label next to the edge, and the numbers in boxes
More informationARTICLE IN PRESS. European Journal of Operational Research xxx (2004) xxx xxx. Discrete Optimization. Nan Kong, Andrew J.
A factor 1 European Journal of Operational Research xxx (00) xxx xxx Discrete Optimization approximation algorithm for twostage stochastic matching problems Nan Kong, Andrew J. Schaefer * Department of
More informationACO Hypercube Framework for Solving a University Course Timetabling Problem
ACO Hypercube Framework for Solving a University Course Timetabling Problem José Miguel Rubio, Franklin Johnson and Broderick Crawford Abstract We present a resolution technique of the University course
More informationGRANULAR MODELLING OF EXAM TO SLOT ALLOCATION Siti Khatijah Nor Abdul Rahim 1, Andrzej Bargiela 1, Rong Qu 2 1 School of Computer Science
GRANULAR MODELLING OF EXAM TO SLOT ALLOCATION Siti Khatijah Nor Abdul Rahim 1, Andrzej Bargiela 1, Rong Qu 2 1 School of Computer Science University of Nottingham, Malaysia Campus Email: abb@cs.nott.ac.uk,
More information