[rules-users] Continuous Planning values for task planning
Josef Bajada
josef.bajada at gmail.com
Tue Sep 11 12:17:57 EDT 2012
Hi Geoffrey,
Thanks for your reply.
Well, my current design is just:
@PlanningEntity class Task
{
@PlanningVariable Task previousTask;
The startTime and endTime then are just derived from the previousTask and
the constraints that come from the facts (i.e. other dependency tasks which
cannot change).
The startTime is computed by taking the max(prevTask.endTime,
max{dependency.endTime + dependency.minoffset}) for each dependency when
calculating the Score, since continuous variables were not that
'recommended' so far.
One thing I wanted to do is 'cache' the startTime of each task once
computed for that Score iteration. This way if A depends on B and B already
had its startTime calculated I don't calculate it again.
On the other hand, to make sure that they don't get carried over from one
iteration to another the first thing the score calculator does is set the
startTime of all planning entities to null.
Not sure if this is the right approach after all, but it was another reason
why I stayed away from the IncrementalScoreCalculator. If I find a more
elegant solution for the startTime calculation (maybe as you are suggesting
with design option 1) it would fit more easily into a Drools drl file...
once I get my head around it.
I like this approach:
@PlanningEntity class Task {
@PlanningVariable Task previousTask;
@PlanningVariable int gap;
}
However, didn't the 'gap' now become a Continuous variable? How do I use
it?
Is it possible to prune the search space in any way so that it only looks
for values between the minOffset and maxOffset of its dependencies?
Re the custom move factory, I'll have a look.
thanks,
Josef
On 11 September 2012 13:33, Geoffrey De Smet <ge0ffrey.spam at gmail.com>wrote:
>
> Op 11-09-12 04:02, Josef Bajada schreef:
>
> Hi Geoffrey,
>
> I managed to implement the solution with the relative end Times. (I
> haven't yet used the incremental score calculation because I haven't quite
> understood how I implement it)
>
> If you use DRL, the incremental score calculation comes automatically.
> Basically, the average calculation score count should be above 1000/s,
> even for cases with 1000+ entities. Without incremental score calcluation,
> the average calculation score count won't be anywhere near that when the
> entityList size scales out.
>
>
> However I am facing one problem with our approach:
>
> Lets say we have 3 tasks A, B, and C, each of duration 1 just for
> simplicity, with these constraints:
>
> a) B can only start between A.endTime + 10 and A.endTime + 100
> b) C can only start between A.endTime + 50 and A.endTime + 100
> c) C can only start between B.endTime + 10 and B.endTime + 20
>
> With our approach of B.startTime = max(prev.endTime, A.endTime + 10) B
> will always be scheduled to start at time 11 (and end at time 12) given the
> hard constraints above.
> On the other hand C cannot start before time 51, thus violating constaint
> C.
>
> Good point. The question is whether you need to go for an implementation
> 1) for which B can have any gap with A between 10 and 100
> or 2) for which B can only chose from a specific set of gap lengths
> between 10 and 100: 11, 99, 51, (51 - B length), ...
> I think 2) is probably not realistic, if you start adding many other tasks
> it becomes unmanageable (unless we are overlooking a clever way to filter
> potential gaps).
>
> Ok, so the question is if we want this design:
>
> 1) @PlanningEntity class Task {
> @PlanningVariable Task previousTask;
> @PlanningVariable int gap;
> }
>
> Or your original design:
>
> 0) @PlanningEntity class Task {
> @PlanningVariable int startTime;
> }
>
> (there might be other designs too)
>
> It kinda depends on your planning window length what might works best:
> Presume you have a planning window length of 100000 and 200 tasks and an
> average maximum gap of 100.
> In case 0), each of 200 tasks will have 1 variable with 100000 values. =>
> that's a search space of 100000^200 = 10^1000.
> In case 1), each of 200 tasks will have 2 variables, one var with 200
> values and one var with on average 100 values. => that's a search space of
> (100 * 200)^200 = 10^860.
> Note that those kind of search space sizes are ok, don't let the numbers
> scare you.
>
>
> The only solution to this would be to move B forward to start at time
> 30. Is there any technique to detect these situations and make such
> adjustments?
>
> In a custom MoveFactory you could detect these situations and generate
> moves for them.
> The advantage of design 1) is that you 'd be effectively pushing all the
> tasks dependent on the task that get moved too.
> If those dependent task don't get pushed back, they 'll cause a hard
> constraint to trigger and the move will be very unlikely to win the local
> search step.
> This is known as a "score trap" (see manual).
>
> You might also want to take a long hard look at as complex as
> BedDesignationPillarPartSwapMoveFactory in the hospital bed planning
> example.
> The idea is to swap 2 chains of tasks, instead of just 2 tasks.
>
>
>
> thanks again,
>
> Josef
>
>
>
>
> On 20 August 2012 09:46, Geoffrey De Smet <ge0ffrey.spam at gmail.com> wrote:
>
>> You 'll want incremental score calculation (with delta's) for your "end
>> times".
>>
>> http://docs.jboss.org/drools/release/5.4.0.Final/drools-planner-docs/html_single/index.html#incrementalScoreCalculation
>>
>> So that naturally puts the calculation of those end times in the scoreDRL
>> (or IncrementalJavaCalculator if you're not using drools).
>> Whether or not that endTime should be a property on the model (at least
>> the model that Planner works with), is an open design question.
>> If it isn't, you can use insertLogicals in DRL, like I did in
>> nurserostering to calculate the number of sequential weekends being worked
>> etc.
>> If it is a property on your model, either the DRL must first "correct it"
>> (with higher salience rules for example),
>> or custom moves must "correct it" as they are being done (which is very
>> hard as it entails constraints functionality).
>>
>> As for the cloning: it's quite simple:
>> Either the model's entity contains the endTime property, then clone it.
>> If it doesn't, then there's nothing to clone.
>> I don't see any problems related to the cloning.
>>
>> Op 18-08-12 01:53, Josef Bajada schreef:
>>
>> Hi Geoffrey,
>>
>> Following your advice and after gaining some more understanding of
>> planner, I think approaching the problem as a chain of tasks one of the
>> other makes sense. It would have some 'wait' time between tasks where the
>> end time of the previous task is smaller than the minimum time the task has
>> to wait after its dependency (which could be different from the previous
>> task).
>>
>> I've noticed that in most examples (TSP and VRP), there is some
>> separation between the model and the planning entity that is being moved
>> around in the chain, which also makes sense. (For instance in TSP, Visit is
>> the planning entity while City is the data entity). When the planning
>> entity gets cloned, the data entity gets assigned to the clone.
>>
>> I am concerned however, that with my dependency between tasks and their
>> end times (which are as such a property of the planning entity not the data
>> entity) I won't be able to model them in this way. (For a task to know
>> whether it has violated its hard constraint it needs to get access to the
>> end time of its dependency, which is in the planning entity). My concern is
>> that I might end up with a whole mess when it comes to the cloning of
>> tasks. I am also concerned about the performance of computing the end time
>> of each node recursively based on its previous task and dependent tasks.
>>
>> What is the best approach in this case?
>>
>> thanks,
>>
>> Josef
>>
>>
>>
>> On 25 July 2012 08:30, Geoffrey De Smet <ge0ffrey.spam at gmail.com> wrote:
>>
>>>
>>> Op 24-07-12 23:14, Josef Bajada schreef:
>>>
>>> Hi Geoffrey,
>>>
>>> Thanks for your reply.
>>>
>>> > Does it make sense to wait longer than 7 mins after task A
>>> (presuming no other task forces occupies the user at that time)?
>>> > Put differently: Can we say that the starting time of B =
>>> Math.max((endTime of task before B), (endTime of task A + 7 minutes))?
>>> > If we can say that, it's pointless to investigate the solutions where
>>> task B starts 8 minutes after task A and the user doing no task that last
>>> minute.
>>>
>>> Yes, the starting time of B = Math.max((endTime of task before B),
>>> (endTime of task A + 7mins)) as long as it is smaller than (endTime of task
>>> A + 8mins).
>>> Yes, it is pointless to investigate the solutions where task B starts 8
>>> minutes after task A and the user doing no task that last minute.
>>> The 8 minute is just a constraint that the task in between tasks A and B
>>> cannot take longer than 7:59s.
>>>
>>> I am thinking that maybe instead of using time itself as the planning
>>> variable, we would use time just to determine the Hard and Soft scores.
>>> So if Task B is scheduled after Task A + 8mins by the solver, then it
>>> inflicts on the hard score. Similarly if Task B is scheduled before Task A
>>> + 7 mins.
>>> Does my reasoning make sense in any way?
>>>
>>> Yes, but personally, I 'd design it differently (although I have no
>>> proof that my way would be better), like this:
>>> "Task B is scheduled after Task A + 8mins by the solver" => make this a
>>> hard constraint
>>> "Task B is scheduled before Task A + 7 mins" => make this a build-in
>>> hard constraint (= not a constraint in the scoreDRL or ScoreCalculator, but
>>> by design, see manual).
>>> Each Task is assigned to a previousTaskOrPerson (and this variable is
>>> chained). It does not know it's startingTime directly.
>>> The scoreDRL or ScoreCalculator calculates the startingTime of a Task
>>> dynamically, by applying this logic:
>>> starting time of B = Math.max((endTime of previousTaskOrPerson of B),
>>> (endTime of task A + 7mins))
>>> Note: "Chained=true" guarantees that there are no cycles of Tasks and
>>> that no Tasks exists with a previousTaskOrPerson == null.
>>> Note: "(endTime of task A + 7mins)" is not hard coded in the score
>>> function: you won't find "7" or "A" in there.
>>>
>>>
>>>
>>> thanks,
>>> Josef
>>>
>>>
>>>
>>> On 24 July 2012 20:46, Geoffrey De Smet <ge0ffrey.spam at gmail.com> wrote:
>>>
>>>>
>>>> Op 23-07-12 16 <23-07-12%2016>:26, Josef Bajada schreef:
>>>>
>>>> Hi Geoffrey,
>>>>
>>>> Well I want to leave 'space' between tasks in the situations where
>>>> there are hard constraints that require me to put this space.
>>>>
>>>> This makes the chaining technique harder to model, but I wouldn't
>>>> write it off yet.
>>>>
>>>>
>>>> As a simple example:
>>>>
>>>> Task A: Put pasta in boiling water (duration 40 seconds)
>>>> Task B: Take pasta out of boiling water (duration 50 seconds, cannot
>>>> start before 7 mins after Task A finishes, cannot start after 8 mins after
>>>> Task A finishes)
>>>>
>>>> Does it make sense to wait longer than 7 mins after task A (presuming
>>>> no other task forces occupies the user at that time)?
>>>> Put differently: Can we say that the starting time of B =
>>>> Math.max((endTime of task before B), (endTime of task A + 7 minutes))?
>>>> If we can say that, it's pointless to investigate the solutions where
>>>> task B starts 8 minutes after task A and the user doing no task that last
>>>> minute.
>>>> If we can say that, then chaining can calculate the the starting time
>>>> of a task on the fly differently.
>>>>
>>>> Task C: Chop vegetables (duration 2 minutes).
>>>>
>>>> This will evidently leave some gaps. The ideal result from the solver
>>>> should be:
>>>>
>>>> Task A: at time 0 (ends at 40s)
>>>> Task C: at time 41s (ends at 2:41)
>>>> Task B: at time 7:40
>>>>
>>>> There is a gap between C and B which is OK.
>>>>
>>>> If another Task is added to the story:
>>>> Task D: Prepare sauce (duration 7 minutes)
>>>>
>>>> I would want the following result:
>>>>
>>>> Task A: at time 0 (ends at 40s)
>>>> Task D: at time 41s (ends 7:41s)
>>>> Task B: at time 8:42s (ends 9:32s)
>>>> Task C: at time 9:33s (ends 11:33s)
>>>>
>>>> Task C can actually take place before Task A too.
>>>>
>>>> I still need to read and understand the chaining functionality
>>>> properly. Do you think it would allow me to achieve the above?
>>>>
>>>> I don't know.
>>>> But using continuous variables in a search problem such as this that
>>>> smells discrete with discrete constraints (A must start before B, ...),
>>>> could blow up the search space unnecessarily.
>>>>
>>>> If you want to look into using continuous variables: the support for it
>>>> is limited currently.
>>>> you can reuse the Drools Planner metaheuristic algorithms (including
>>>> termination, score, ...), but there's no decent generic move factory
>>>> support for continuous variables yet.
>>>> So you 'll have to write a custom MoveFactory that creates a limited
>>>> subset of moves.
>>>> Also, construction heuristics can't handle continuous variables yet, so
>>>> you 'll have to write a custom SolutionIntializer.
>>>> There are examples with a custom MoveFactory and a custom
>>>> SolutionIntializer where you can copy paste from, but none with continuous
>>>> variables at the moment.
>>>>
>>>> thanks,
>>>>
>>>> Josef
>>>>
>>>>
>>>>
>>>> On 22 July 2012 20:05, Geoffrey De Smet <ge0ffrey.spam at gmail.com>wrote:
>>>>
>>>>> Presuming that you don't want to leave space between tasks, you can
>>>>> design your model differently by using the "chained" functionality:
>>>>> it will be far more efficient and the planning variable won't be
>>>>> continuous.
>>>>>
>>>>> Let's presume you're scheduling Tasks to Persons.
>>>>>
>>>>> @PlanningEntity
>>>>> class Task implements TaskOrPerson {
>>>>>
>>>>> ...
>>>>>
>>>>> @PlanningVariable(chained = true)
>>>>> @ValueRanges({
>>>>> @ValueRange(type = ValueRangeType.FROM_SOLUTION_PROPERTY,
>>>>> solutionProperty = "taskList"),
>>>>> @ValueRange(type = ValueRangeType.FROM_SOLUTION_PROPERTY,
>>>>> solutionProperty = "personList",
>>>>> excludeUninitializedPlanningEntity = true)})
>>>>> public TaskOrPerson getPreviousTaskOrPerson() {
>>>>> return previousTaskOrPerson;
>>>>> }
>>>>>
>>>>> public int getDuration() {
>>>>> return duration;
>>>>> }
>>>>>
>>>>> public int getStartingTime() {
>>>>> int startingTime = 0;
>>>>> TaskOrPerson taskOrPerson = getPreviousTaskOrPerson();
>>>>> while (taskOrPerson instanceof Task) { // Every chain is
>>>>> guarantee to end up with an anchor (= Person)
>>>>> startingTime += ((Task) taskOrPerson).getDuration();
>>>>> taskOrPerson = ((Task)
>>>>> taskOrPerson).getPreviousTaskOrPerson()
>>>>> }
>>>>> return startingTime;
>>>>> }
>>>>>
>>>>> }
>>>>>
>>>>> class Person implements TaskOrPerson {
>>>>>
>>>>> }
>>>>>
>>>>> For a good example, take a look at the VehicleRouting example.
>>>>> For more info about chaining, in the manual see section 4.3.4.2.6.
>>>>> Chained
>>>>>
>>>>> http://docs.jboss.org/drools/release/5.4.0.Final/drools-planner-docs/html_single/index.html
>>>>>
>>>>> Op 22-07-12 18 <22-07-12%2018>:00, Josef Bajada schreef:
>>>>>
>>>>> Hi,
>>>>>
>>>>> I am new to Drools and Drools Planner, so apologies if I am asking
>>>>> anything obvious.
>>>>>
>>>>> My objective is to implement a simple (for now) planner which
>>>>> schedules tasks according to 2 main criteria:
>>>>> - Their duration (in seconds)
>>>>> - Their dependencies on other tasks (e.g. Hard Constraint that Task B
>>>>> has to start between 180 and 200 seconds after Task A finishes).
>>>>>
>>>>> Since there are gaps between dependent tasks as part of the hard
>>>>> constraints other tasks can be fitted in between dependent tasks.
>>>>> So the Solver needs to find the optimal start time for each task that
>>>>> satisfies the hard constraints, and in the shortest total timeline possible
>>>>> to complete all tasks (soft constraint).
>>>>>
>>>>> The main problem I am finding is that this start time, which is
>>>>> essentially the planning variable is a continuous variable.
>>>>> Chapter 4 of the Drools documentation mentions very briefly (Section
>>>>> 4.3.4.1) that planning variables can be continuous, but there does not
>>>>> seem to be any more details about how to achieve this.
>>>>>
>>>>> Even if the planning variable was discrete (say bins of 5 second
>>>>> intervals), there is no upper bound as such.
>>>>>
>>>>> How is it best to handle such planning variables in Drools Planner?
>>>>>
>>>>> thanks,
>>>>> josef
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> rules-users mailing listrules-users at lists.jboss.orghttps://lists.jboss.org/mailman/listinfo/rules-users
>>>>>
>>>>>
>>>>> --
>>>>> With kind regards,
>>>>> Geoffrey De Smet
>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> rules-users mailing list
>>>>> rules-users at lists.jboss.org
>>>>> https://lists.jboss.org/mailman/listinfo/rules-users
>>>>>
>>>>>
>>>>
>>>>
>>>> _______________________________________________
>>>> rules-users mailing listrules-users at lists.jboss.orghttps://lists.jboss.org/mailman/listinfo/rules-users
>>>>
>>>>
>>>> --
>>>> With kind regards,
>>>> Geoffrey De Smet
>>>>
>>>>
>>>> _______________________________________________
>>>> rules-users mailing list
>>>> rules-users at lists.jboss.org
>>>> https://lists.jboss.org/mailman/listinfo/rules-users
>>>>
>>>>
>>>
>>>
>>> _______________________________________________
>>> rules-users mailing listrules-users at lists.jboss.orghttps://lists.jboss.org/mailman/listinfo/rules-users
>>>
>>>
>>> --
>>> With kind regards,
>>> Geoffrey De Smet
>>>
>>>
>>> _______________________________________________
>>> rules-users mailing list
>>> rules-users at lists.jboss.org
>>> https://lists.jboss.org/mailman/listinfo/rules-users
>>>
>>>
>>
>>
>> _______________________________________________
>> rules-users mailing listrules-users at lists.jboss.orghttps://lists.jboss.org/mailman/listinfo/rules-users
>>
>>
>> --
>> With kind regards,
>> Geoffrey De Smet
>>
>>
>> _______________________________________________
>> rules-users mailing list
>> rules-users at lists.jboss.org
>> https://lists.jboss.org/mailman/listinfo/rules-users
>>
>>
>
>
> _______________________________________________
> rules-users mailing listrules-users at lists.jboss.orghttps://lists.jboss.org/mailman/listinfo/rules-users
>
>
>
> _______________________________________________
> rules-users mailing list
> rules-users at lists.jboss.org
> https://lists.jboss.org/mailman/listinfo/rules-users
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.jboss.org/pipermail/rules-users/attachments/20120911/8188d37d/attachment-0001.html
More information about the rules-users
mailing list