US 5,319,781 AGrant
Generation of Schedules Using a Genetic Procedure
Issue Date:1994-06-07
•14 Claims
•5 Drawing Sheets
Abstract
In the scheduling method disclosed herein, a genetic algorithm is employed to improve a population of possible schedules represented by respective chromosomes, where the chromosomes upon which the genetic algorithm operates are not a direct encoding of a possible schedules. Rather, the details of the scheduling problem and the real life constraints typically associated with such problems are hidden from the genetic algorithm by the use of a deterministic schedule builder which operates on lists of the desired tasks and which generates legal schedules, i.e. schedules which do not violate hard constraints. The legal schedules so generated are evaluated or scored and the scores are provided to the genetic algorithm as feedback for influencing subsequent operation of the genetic algorithm.
Metadata
Assignee
- Bolt Beranek and Newman Inc.
Inventor
- Gilbert P. Syswerda
Application Information
Application Number:US 6953115
Filing Date:1991-05-03
Priority Date:1991-05-03
Art Unit:236
Classifications
IPC:
G06F 1500G06F 900
Field of Search:
395364650DIG. 1;281.3;281.8
Patent Drawings (5 sheets)
Description
Background of the Invention
The present invention relates to a method of scheduling and more particularly to a method of scheduling tasks which employs a genetic algorithm for progressively improving possible solutions to a scheduling problem.
As is understood by those skilled in the art, scheduling and planning are difficult problems which do not easily yield to automated solving. They are problems which are typically tackled with a combination of search techniques and heuristics. Scheduling is difficult for various reasons. It is a computationally complex problem described in computer science terms as NP complete. This means that search techniques that deterministically and exhaustively search the space of possibilities will probably fail because of time requirements. On the other hand, search techniques that use heuristics to prune the search space will not be guaranteed to find an optimal solution or even a good one. Further, scheduling problems are often complicated by the details of a particular scheduling task. Algorithmic consideration of the specific constraints must often be embodied in what amounts to a domain specific expert system. Thus, any given such system has limited applicability. How the expert knowledge is embedded into an optimization algorithm is often very algorithm specific.
It has previously been proposed to apply genetic algorithms to scheduling problems. As is understood by those skilled in the art, genetic algorithms function by applying operators, e.g. crossover and mutation, to a population of possible solutions often referred to in the art as "chromosomes". Traditionally, chromosomes are simple binary vectors. This simple representation has substantial appeal and the theoretical grounding of genetic algorithms is based on the binary vectors and simple operators. However, many problems emerge when one attempts to represent complicated problems using binary vectors. The success of genetic algorithms in finding good solutions is in general rather directly a function of the manner in which the chromosomes are encoded to represent possible solutions. The situation is further complicated in scheduling problems in that the usual methods of applying crossover and mutation may generate illegal solutions and the genetic algorithm will then in effect spend substantial effort in exploring unhelpful search space. Encoding methods attempting to avoid these problems are described in Davis, L. (1985). Job shop scheduling with genetic algorithms. Proceedings of an International Conference on Genetic Algorithms and Their Applications. Hillsdale, N.J.: Lawrence Erlbaum Associates. and Whitley, D., T. Starkweather, and D. Fuquay (1989). Scheduling problems and the traveling salesman: the genetic edge recombination operator. Proceedings of the Third International Conference on Genetic Algorithms and Their Applications. San Mateo, Calif.: Morgan Kaufmann.
Among the several objects of the present invention may be noted the provision of a novel method for creating schedules; the provision of such a method which will effectively schedule tasks, each of which requires selected resources from a predetermined pool of resources for respective time intervals; the provision of such a method which will schedule tasks having associated therewith various constraints, some of which may be hard or rigid and others of which may be more in the nature of preferences; the provision of such a method which will consistently generate good solutions to complex scheduling problems; the provision of such a method which is easily adaptable to a variety of scheduling problems; the provision of such a method which is adapted to automated or computerized implementation; the provision of such a method which operates quickly; and the provision of such a method which is highly reliable and which may be implemented relatively inexpensively. Other objects and features will be in part apparent and in part pointed out hereinafter.
In accordance with the practice of the present invention, a genetic algorithm is employed which utilizes chromosomes which are not simple binary vectors but, rather, are more indirect representations of possible solutions, for example, merely ordered lists of tasks to be performed. Further, the genetic algorithm is insulated from domain specific knowledge by the inclusion, in the overall system, of a deterministic schedule builder. The schedule builder operates to place the tasks in a given list into a legal schedule, i.e. a schedule which complies with the various hard constraints which may be associated with each task. Legal schedules generated in this manner are then evaluated or scored on the basis of the various soft or preferential constraints to determine how good the putative schedule is. The scores of possible solutions are provided as feedback to the genetic algorithm for use in controlling the operators which effect evolution of the population of solutions.
In a preferred embodiment, the method of the present invention operates to schedule tasks, each of which requires selected resources from a predetermined pool of resources for respective time intervals, each task having associated constraints and an associated scoring function for evaluating its inclusion within a schedule. An initial population of sequential task lists is generated, preferably randomly. The successive tasks in each list are sequentially fit into a respective legal schedule, observing the associated constraints. Each possible schedule in the population is evaluated in accordance with the scoring function to obtain a ranking of the respective lists. Pairs of higher ranking lists are stochastically selected for combination to generate new member lists to be added to the population and lower ranked lists are stochastically deleted from the population. Preferably, mutation is periodically applied by stochastically selecting a pair of tasks in one of the task lists and reversing the positions of the pair in that list.
Brief Description of the Drawings
FIG. 1A is a matrix illustrating an exemplary set of tasks to be scheduled;
FIG. 1B is a partially constructed schedule placing some of the tasks of FIG. 1A into a possible schedule;
FIG. 2 is a diagrammatic illustration of a schedule generating system operating in accordance with the present invention;
FIG. 3 is a diagram illustrating a crossover operator employed in the genetic algorithm component of the system of FIG. 2;
FIG. 4 is a diagrammatic illustration of another type of crossover operator;
FIG. 5 is a flowchart illustrating the operation of the system of the present invention;
FIG. 6 is a graph illustrating the operation of one embodiment of the invention operating on a model scheduling problem; and
FIG. 7 is a flowchart illustrating an alternate form of a schedule building component for use in a system of the present invention.
Corresponding reference characters indicate corresponding parts throughout the several views of the drawings.
Description of the Preferred Embodiments
The complex nature of the scheduling problem may be illustrated by the simplified example illustrated in FIGS. 1A and 1B. FIG. 1A illustrates a set of tasks labeled a-t down the left side of the matrix together with the number of hours over which various resources will be required. Thirty different resources are indicated by different positions along the horizontal axis of the matrix of FIG. 1A. There is only a single instance of each resource so a conflict for any resource will be a violation of a hard constraint. Other examples of hard constraints may be that no task can be scheduled after certain times of the day. Likewise, particular users may not be available at particular times or may prefer a particular time. There may in some situations also be constraints of precedence, i.e. certain tasks cannot be performed until other tasks have been completed.
FIG. 1B is a partially constructed schedule with the tasks a-g placed into the schedule on a simple first-come, first-serve basis but observing the hard constraint of avoiding conflict in the use of a resource. A schedule builder which functions in this way is deterministic in that the schedule which results is determined by the order in which the tasks occur in the list. More complex schedule builders may also be implemented which utilize more domain specific knowledge, such as soft constraints, but which are still deterministic in the sense that the resulting schedule is determined by the order in which the tasks are considered.
The method of the present invention may be conveniently understood as comprising three portions. With reference to FIG. 2, these three portions are a deterministic schedule builder as described above, designated by reference character 11; a schedule evaluator, designated by reference character 13; and a genetic algorithm 15. These portions or components operate on a population of possible solutions 17 maintained by the genetic algorithm.
In the genetic algorithm component of the system of the present invention, operators work on a population of members or chromosomes which are not simple binary vectors but, rather, are lists of tasks in various permutations of order. As explained in greater detail hereinafter, the nature of the operators used is such that the lists which constitute the population are always complete, that is, they always include all the tasks to be considered. As indicated previously, there will be associated with each of the tasks certain preferences or soft constraints. These preferences or soft constraints provide a mechanism for evaluating or scoring the schedules which are generated from the task lists by the deterministic schedule builder. Clearly, the scoring or evaluation will be highly dependent upon the specific nature of the scheduling task, that is, the evaluation function must capture what it is about schedules that make them seem good or bad to the users of the system. For example, there will typically be a priority associated with each task veering from preemptory to casual. Further, when not all requested tasks can be placed into a schedule, those with the lowest relative priority should be left out and the scoring function should take this into consideration. While the preferred type of genetic algorithm described herein is of the general type described as a "steady state" genetic algorithm, it should be understood that the method of the present invention could also be utilized with a genetic algorithm of the more common "generational" type in which entire populations are repetitively replaced.
While the preferred embodiment described herein is described in terms of scheduling tasks so as to best utilize available resources, it should be understood that the techniques of the present invention may also be applied to the scheduling or deploying of resources to tasks or events which are essentially predetermined in time in which case the chromosomes operated on by the genetic algorithm may be lists of the resources.
In a model scheduling problem used for the testing of the method of the present invention, the method was utilized to construct possible schedules for 90 tasks, each task requiring one to three hours and a set of up to 30 resources, i.e. this was a larger set similar to the example of FIG. 1. The 90 tasks were to be scheduled into a block of 40 hours. There was only a single instance of each resource so that a conflict for any resource was considered a hard constraint violation. Tasks were given respective priority values ranging from 1 to 100. A further soft constraint was added to specify that some of the tasks preferred to be scheduled during the first 12 hours.
For the model problem which was used to establish the performance characteristics described in greater detail hereinafter, the evaluation of schedules generated by the schedule builder was performed as follows. The evaluation of a schedule started with the sum of the priorities of all the tasks. For each task not placed into the schedule, its priority was subtracted from the sum. For each task placed into the schedule, its priority was added to the sum unless there was a soft constraint violation in which case only half of its priority was added. With this evaluation scheme, if no task were scheduled, the evaluation function returns zero. If a perfect schedule were constructed, the evaluation function would return twice the priority sum. In the model problem, however, construction of a perfect schedule was not possible since the number of tasks was too great for the time allotted.
In the genetic algorithm component of the method of the present invention, an initial population of such chromosomes is generated either entirely randomly or by some scheme employing both a random component and a domain specific knowledge. As is understood in the art, slanting or skewing of initial populations by the use of domain specific knowledge can speed up at least the initial operation of a genetic algorithm but it may cause the algorithm to miss possibly superior solutions.
In order to cause the population of possible solutions to evolve, the genetic algorithm component of the method of the present invention preferably employs a form of both crossover and mutation operators, the completeness of the member lists being preserved. As in known genetic algorithm systems, members of the existing population to which the operators are applied are selected stochastically based upon their score or ranking within the population. By the term "stochastically" is meant that a member with a relatively high score has a higher likelihood of being chosen or selected for participation in the operation than a relatively low scoring member but that there is also a randomizing component so that merely having a high score does not assure selection.
The preferred crossover operator used in the method of the present invention may conveniently be designated a position based crossover. For the crossover operation, two members of the existing population of lists are selected stochastically as described previously. The two selected members are conventionally referred to as "parents". A set of positions is selected randomly. At each selected position, the position of the corresponding task in one parent are imposed on the other parent. The other tasks in the list are shifted to accommodate the moved tasks so that the list is maintained complete. This imposition of task position from one parent on another can be performed in both directions so that two new members or lists are generated which can be added to the population. This position based crossover is illustrated in FIG. 3. Crossover operations of this type are referred to in the art as combination or recombination operators.
Another crossover operator effective in the method of the present invention may conveniently be termed order based crossover. Again, a set of positions is randomly selected. The order of tasks in the selected positions in one parent is imposed on the corresponding tasks in the other parent. Again, tasks in other positions in the list are shifted to accommodate the changes and maintain a list which is complete. This form of order based crossover is illustrated in FIG. 4.
The method of the present invention also contemplates that a mutation operator will be applied at intervals to a member of the existing population. Again, the list or member of the existing population selected for the operation is chosen stochastically. The preferred form of mutation operator may be designated order based mutation. In this operation, two tasks in the chosen list are selected at random and their positions are interchanged.
Alternate but less effective mutation operators which have been used may be termed position based mutation and scramble mutation. In position based mutation, two tasks in the chosen list or member chromosome are selected at random and the second task is placed before the first. In the scramble mutation operator, a set of positions within the list is chosen randomly and the order of the tasks in the selected positions are scrambled randomly.
As indicated previously, the method of the present invention has, for evaluation, been applied to a model problem which rather closely simulates a real world situation. The application of the method to this problem can be described with reference to the flowchart of FIG. 5 which illustrates the operation of the method, with further reference to actual numerical examples. The problem involved 90 tasks, each requiring one to three hours and a set of up to 30 resources. The 90 tasks were to be scheduled into a block of 40 hours.
With reference to FIG. 5, the initial step, designated by reference character 41, is to generate an initial population of task lists. In the tests described hereinafter, the population comprised 30 members and each was a complete list including all of the tasks. In the tests described hereinafter, the initial population in each case was generated entirely randomly.
As indicated in block 43, each list in the initial population was then submitted to the deterministic schedule builder to generate a legal schedule, that is, a schedule which did not violate any hard constraints. In the test described in detail hereinafter, each successive task was fit into the schedule on a first come, first serve basis and was put into the schedule at the first position which would legally accept the task. A more complex schedule builder is described hereinafter but was not utilized in the numerical examples included in this specification. As indicated at block 45, each schedule so generated was then scored or evaluated using the evaluation function described previously herein.
As indicated previously, it is preferred that both crossover and mutation operators be utilized in the practice of the present invention. The respective rate or use of each operator can be either fixed or variable over the course of the operation of the genetic algorithm. A selection of either crossover or mutation to be employed on each cycle is indicated at block 47. In the case of constant and equal application of the two types of operators, this selection is simply an alternation between the two types. The mutation and crossover operations are designated by reference characters 51 and 53, respectively. The selected operators act as described previously herein to generate one or two new members. Again, the new members are complete lists of tasks. The new member lists are applied to the schedule builder and evaluated as indicated at block 57 and, if they do not already exist in the population, the new members are added to the population.
In order to maintain the population at the desired level, i.e. 30 members, lower scoring members are stochastically deleted from the population as indicated at block 58. The process then loops back to block 47 until a condition for completion of the running of the genetic algorithm is met. In the tests described herein, the genetic algorithm was run until a predetermined number of evaluations had been performed, i.e. 3000 evaluations. At this point, the current population was evaluated for its highest scoring member as indicated at block 59 and the corresponding schedule was output by the system. As should be understood, it will typically be appropriate to output several of the better scoring members for evaluation by a user's subjective standards.
The performance of the genethic algorithm in increasing the scoring of the best member in a population of lists which was initially created randomly is illustrated in FIG. 6. Curve 71 illustrates the progressive improvement obtained by using fixed and equal levels of position based crossover and order based mutation. Curve 73 illustrates the results using sliding probabilities for the application of each operator. In the sliding probability tests, the order based mutation rate was slowly increased while the position based crossover rate was gradually decreased during the running of the algorithm. As may be seen, this produced a slight improvement over the fixed rate tests. The basis for the curves plotted is the score of the best member in the population at any given point in the run. Further, 50 runs were made for each type and the values obtained were averaged for each corresponding point in each run to obtain the values used for plotting the curves 71 and 73.
To provide a frame of reference for illustrating that the application of the genetic algorithm component in the system of the present invention provided significant improvement, FIG. 6 also includes a curve, designated by reference character 76, which illustrates the gradual improvement of score which would be obtained if new task lists were just repetitively and continuously generated randomly, the lists being submitted to the schedule builder and the resulting schedules being scored in the same manner. Since the schedule builder itself will always attempt to schedule as many tasks as possible, the resultant scores are not orders of magnitude below those generated by the complete system and, understandably, show a gradual continuing increase over time. This is also in some ways a reflection of the manner of scoring employed.
Since the initial populations for each of the 50 runs which were averaged to arrive at curve 73 are generated and randomly and separately, it will be understood that some runs will be markedly better than others at any given point during the run and that curve 73 just presents the average at each point. Curve 77 represents the highest score of the then best run of the 50 runs which were done using sliding probabilities. The chance of occurrences such as this suggest that, in certain circumstances, it may be better to do a series of short runs rather than continuing a single run for an equivalent length of time. A demonstration program which operated on the model program as described herein is included as an appendix to this specification. The program is written in the Pascal programming language.
As described hereinbefore, the schedule builder employed in testing with the model problem merely inserted each task into the schedule at the first position where it would legally fit. As indicated previously, it is possible to incorporate more intelligence or domain specific knowledge into the schedule PG,17 builder to cause it to build better schedules, though still acting deterministically as a function of any given task list presented to it. FIG. 7 illustrates one such modification of the schedule builder.
The procedure illustrated in FIG. 7 also considers each task in the list in succession as indicated at block 81 but tries a preselected number of legal positions for inserting each successive task into the partial schedule built thus far. After each possible position is determined as indicated at block 83, the partial schedule resulting is then evaluated as a function of the soft or preferential constraints -- block 85. The number of positions (N) for which the schedule builder evaluates a given task in the current list may be considered to be the depth of local search. After the preselected number of possibilities are tried, as tested at block 87, the best of the set is selected for addition into the new partial schedule before passing on to the next task in the current member list. In effect, the looping procedure of FIG. 7 can be substituted for the single step procedure indicated at block 43 in FIG. 5.
It is not, however, deemed desirable to add substantial intelligence to the schedule builder, i.e. to create a machine which in itself nearly performs a heuristic solving of the scheduling problem. If the schedule builder is made smarter, more and more chromosomes will result in the same schedule and thus the same evaluation. This will hide information from the genetic algorithm since a chromosome and a somewhat worse chromosome can both result in the same evaluation. In other words, if the schedule builder is not allowed to generate bad schedules in response to correspondingly bad list orderings of the tasks, the genetic algorithm component of the system will not have the necessary flexibility and may not consider certain arrangements which could lead to very good schedules. Rather, the method of the present invention is believed to perform best if the genetic algorithm is insulated from the domain specific knowledge as much as possible, the schedule builder operates to simply build legal schedules observing hard constraints, and the ranking of possible solutions is left to the schedule evaluator which takes into consideration the soft or preferential constraints.
While the operation of the system of the present invention has been described in the context of making all placements of the tasks to a schedule, it should be understood that the method may also be applied to a partially filled-in schedule. For example, if the user may have to place certain tasks at certain locations within the schedule for reasons which are sufficient to the user, it may not easily be encoded into the process. Such usage should also be understood to fall within the scope of the present invention.
In view of the foregoing it may be seen that several objects of the present invention are achieved and other advantageous results have been attained.
As various changes could be made in the above constructions without departing from the scope of the invention, it should be understood that all matter contained in the above description or shown in the accompanying drawings shall be interpreted as illustrative and not in a limiting sense. ##SPC1##
Claims
What is claimed is:
1. A computer implemented method of scheduling tasks each of which has associated hard and soft constraints, said method comprising: generating an initial population of sequential task lists; sequentially fitting the tasks in each list in the then existing population in order into a respective possible schedule observing associated hard constraints thereby to deterministically generate a set of schedules; evaluating each possible schedule so generated as a function of the associated soft constraints to obtain a ranking of the respective lists; repetitively applying to the then existing population a genetic algorithm which effects recombination and mutation to members of the population stochastically chosen on the basis of the ranking of the corresponding schedules, thereby to modify the population; and outputting the schedules generated by the last population, together with their rankings.
2. A computer implemented method of scheduling tasks each of which has associated hard and soft constraints, said method comprising: generating an initial population of sequential task lists; sequentially fitting the tasks in each list in the then existing population in order into a respective possible schedule observing associated hard constraints thereby to deterministically generate a set of schedules; evaluating each possible schedule so generated as a function of the associated soft constraints to obtain a ranking of the respective lists; repetitively performing the steps of: selecting a pair of higher ranking lists for combination; imposing characteristics from one of the lists in each selected pair onto corresponding portions of a copy of the other list in the respective pair, the non-corresponding portions of said copy being shifted to accommodate the imposed portions so that the list remains complete and adding the thereby modified copy to the then existing population as a new member; sequentially fitting the tasks in each new member list in order into a respective schedule observing associated hard constraints and evaluating each schedule so generated to obtain a new ranking of the population; deleting a lower ranked list from the then existing population; and outputting the schedules generated by the last population, together with their rankings.
3. The method as set forth in claim 2 wherein said imposing of characteristics is performed by randomly selecting list positions and imposing the corresponding tasks from one of the lists in the pair which occur at those positions onto the corresponding positions in the other list of the pair.
4. The method as set forth in claim 2 wherein said imposing of characteristics is performed by randomly selecting list positions and imposing the order of the tasks from one of the lists in the pair which occur at those positions onto the corresponding tasks in the other list of the pair.
5. The method as set forth in claim 2 further comprising: at intervals, selecting one of the task lists in the existing population and mutating the positions of tasks in that list.
6. The method as set forth in claim 5 wherein said mutation is performed by randomly selecting a pair of tasks in the selected list and reversing the positions of the selected tasks.
7. The method as set forth in claim 2 wherein said modified copy is added to the population only if it does not duplicate an existing member of the population.
8. A computer implemented method of scheduling tasks each of which require selected resources from a predetermined pool of resources for respective time intervals and each of which has associated constraints and an associated scoring function for evaluating its inclusion within a schedule, said method comprising: generating an initial random population of sequential task lists; sequentially fitting the tasks in each list in the existing population in order into a respective possible schedule observing said associated constraints thereby to deterministically generate a set of schedules; evaluating each possible schedule in accordance with said scoring functions to obtain a ranking of the respective lists; repetitively performing the steps of: stochastically selecting a pair of higher ranking lists for combination; randomly selecting tasks from one of the lists in each selected pair and imposing the list positions of those selected tasks on the corresponding tasks in a copy of the other list in the respective pair, the non-corresponding tasks in said copy being shifted in list position to accommodate the imposed positions so that the list remains complete and adding the thereby modified copy to the existing population as a new member; and sequentially fitting the tasks in each new member list into a respective schedule observing associated constraints thereby to deterministically generate a set of schedules and evaluating each schedule so generated to obtain a new ranking of the then existing population; and stochastically deleting a lower ranked list from the then existing population; at intervals, stochastically selecting a pair of tasks in one of the task lists in the existing population and reversing the positions of the selected tasks in that list; and outputting the schedules generated by the last population, together with their rankings.
9. The method as set forth in claim 8 further comprising: at intervals, selecting one of the task lists in the existing population and mutating the positions of tasks in that list.
10. A computer implemented method of assigning resources to service tasks each of which require selected resources for respective time intervals, there being associated hard constraints which must be observed and associated soft constraints for evaluating a schedule, said method comprising: generating an initial random population of sequential lists of items which may comprise tasks or resources; deterministically fitting the sequential member items in each list in the existing population in order into a respective possible schedule observing said associated hard constraints thereby to deterministically generate a set of schedules; evaluating each possible schedule in accordance with said soft constraints to obtain a ranking of the respective lists; repetitively performing the steps of: stochastically selecting a pair of higher ranking lists for combination; imposing characteristics from one of the lists in each selected pair onto corresponding portions of a copy of the other list in the respective pair, the non-corresponding portions of said copy being shifted in to accommodate the imposed portions and adding the thereby modified copy to the existing population as a new member; deterministically fitting the items in each new member list into a respective schedule observing hard constraints and evaluating each schedule so generated to obtain a new ranking of the population; and stochastically deleting a lower ranked list from the existing population; at intervals, selecting a pair of tasks in one of the task lists in the existing population and mutating the positions of tasks in that list; and outputting the schedules generated by the last population, together with their rankings.
11. The method as set forth in claim 9 wherein said imposing of characteristics is performed by randomly selecting list positions and imposing the corresponding items from one of the lists in the pair onto the corresponding positions in the other list of the pair and wherein said mutation is performed by reversing the positions of a randomly selected pair of items in the selected list.
12. The computer implemented method of assigning resources to service tasks each of which require selected resources for respective time intervals, there being associated hard constraints which must be observed and associated soft constraints for evaluating a schedule, said method comprising: generating an initial random population of sequential lists of items which may comprise tasks or resources; deterministically fitting the sequential member items in each list in the existing population in order into a respective possible schedule observing said associated hard constraints thereby to deterministically generate a set of schedules; evaluating each possible schedule in accordance with said soft constraints to obtain a ranking of the respective lists; repetitively applying to the then existing population a genetic algorithm which effects recombination and mutation to members of the population stochastically chosen on the basis of the ranking of the corresponding schedules thereby to modify the population; and outputting the schedules generated by the last population, together with their rankings.
13. A computer implemented schedule generating system which operates on a population of lists of tasks each of which has associated hard and soft constraints, said system comprising: a schedule builder which operates on each list in an existing population to deterministically create a possible schedule which observes associated hard constraints; a schedule evaluator which scores each possible schedule generated as a function of associated soft constraints; and a genetic algorithm which operates to repetitively; stochastically select a pair of higher ranking lists from the then existing population for combination; randomly select tasks from one of the lists in each selected pair and imposing the list positions of those selected tasks on the corresponding tasks in a copy of the other list in the respective pair, the non-corresponding tasks in said copy being shifted in list position to accommodate the imposed positions so that the list remains complete; adding the thereby modified copy to the existing population with the list being submitted to the schedule builder and schedule evaluator so as to obtain a new ranking of the population; stochastically deleting a lower ranked list from the existing population; and at intervals, randomly selecting a pair of tasks in one of the task lists in the existing population and reversing the positions of the selected tasks in that list; and outputting the schedules generated by the last population, together with their rankings.
14. The schedule generating system as set forth in claim 13 wherein the copy is added to the population only if it does not duplicate an existing member of the population.
Patent Citations (1)
| Patent | Date | Inventor | Cited By |
|---|---|---|---|
| US5148513 | 1992-09-01 | Koza et al. |
Non-Patent Literature (2)
- "A Genetic Algorithm for job shop", Falkenauer, E., and Bouffouix, S., Proc. of 1991 IEEE Int. Conf. on Robotics & Aut. Sacramento, Calif. Apr. 1991, pp. 824-829.
- "Efficient Multiprocessor Scheduling Based on Genetic Algorithms", Hou, E. S., Hong, R. and Ansari, N. pp. 1239-1243.