1 / 38

Constraint-Based Scheduling in the Real World

Constraint-Based Scheduling in the Real World. Mark Boddy Honeywell Laboratories Mark.Boddy@honeywell.com. Outline. Crude oil scheduling demo (first example) Discussion of scheduling problems Constraint-based scheduling Avionics scheduling (second example)

mccoolm
Download Presentation

Constraint-Based Scheduling in the Real World

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Constraint-Based Scheduling in the Real World Mark Boddy Honeywell Laboratories Mark.Boddy@honeywell.com

  2. Outline • Crude oil scheduling demo (first example) • Discussion of scheduling problems • Constraint-based scheduling • Avionics scheduling (second example) • Some (very) open questions and irresponsible remarks

  3. Finite-Capacity Scheduling Specify activities linked to operations (transfers, tool changes, delivery dates, production runs, etc.). This specification includes: • Activity start and end times • Resource usage, including raw materials, power, equipment needed, labor required Finite capacity schedules can be used to determine predicted inventory levels, campaign completion/product delivery dates, economic performance, unit operating modes, and resource bottlenecks.

  4. Complex Operational Domains Logistics Discrete Mfg Continuous Mfg Air Transportation Satellite Operations Autonomous Vehicles

  5. Crude Oil Scheduling Demo

  6. Another view of the problem Product Volume Time

  7. Value of Scheduling for Manufacturing • Reduce Costs • Reduce disruption frequency and severity (what-if scenarios, detailed maintenance schedules) • Improve parts/raw material ordering (better tracking of on-hand inventory) • Reduce Work-In-Process inventory (reduced lot-rot, more timely vendor ordering, better synchronization of production to delivery dates) • Reduce changeover costs (cleaning, new catalyst, temperature changes, …). • Enhance Capital Investment Planning • prioritize and quantify payoff for capital investment • Increase Effective Capacity • Identify and remove bottlenecks • Increase agility (improve “Available to Promise”)

  8. What is Scheduling? Determining when and how to accomplish some set of tasks. The classic job shop: • A set J of jobs to be run, each with • a set of tasks to run in sequence, each on some unique element of • a set M of machines Common problem statements: • Find a feasible schedule. • Minimize makespan. • Given a set of deadlines, minimize tardiness.

  9. Possible Complications (a brief sample) • Choice of resource • Other resource constraints • (Global) capacity constraints • Inter-activity constraints • Consumable resources • Complex temporal constraints (latency, preemption, hierarchical activity relationships). • System dynamics (flow rates, chemical composition, …) • Activity generation • Reasoning about state

  10. Manufacturing IT Architecture Strategic Planning Rough-cut Scheduling (Bill of Materials) Production Planning Distribution Receiving Vendor orders Finite-capacity Plant scheduling Capital investment, Finance, Long-term supply and demand forecasts More than 6 months Rough-cut capacity planning, Orders forecasts, Long-leadtime vendor orders 1 month to 1 year MRP-II 2 weeks to 3 months Transportation and Warehouse Management 1 day to 1 month Manufacturing Execution 1 hour to 1 day

  11. How schedules are used is a factor as well • Schedules are generated to satisfy existing plans. • Schedules are generated in a historical context. • Schedules are usually not constructed completely automatically. • Schedules frequently require input from multiple parties. • Schedules are used in operations. • Updating, as events occur. • Rescheduling, as conflicts are detected. • Post mortem examinations. • (Schedules are persistent artifacts.) • Schedules are large. • Optimization is important, but hard to formalize properly.

  12. Scheduling as a Constraint Satisfaction Problem CSPs are specified as: • A set V of variables • A set C of constraints, each constraint a relation specifying tuples of permissible values for some subset of V. The objective is to find a feasible (alt., optimal) complete assignment for V, consistent with C.

  13. Advantages to a CSP approach • Flexible, declarative representation. • Lots of previous and current work on solution methods. • General solution techniques: • Static structural analysis • Propagation • Search • Requirements and scheduling decisions can be represented as constraints.

  14. Scheduling as Search Using a Dynamic CSP CSP Variables: • Activity start and end times • Activity resource assignments Constraints: • Temporal constraints (duration, ordering, release times, deadlines, …) • Resource constraints (permissible assignments, usage requirements, state information, ...) • System dynamics (rate limitations, allowable state changes, …) Search over: • Activity generation • Resource assignments • Activity orderings, start and end times

  15. Hybrid Systems Hybrid systems have both continuous and discrete components, as opposed to: • Combinatorial problems, such as knapsack, TSP, or jobshop • Continuous problems, captured as sets of mathematical equations and inequalities. Combinatorial, continuous, and hybrid constraint problems can all be framed in terms of either satisfiability (CSP) or optimality (COP). Scheduling is a hybrid constraint problem.

  16. Constraint Envelope Scheduling Search variables: • Unary resources: < A before B, B before A > • Capacity resources: < A before B, B before A, A overlaps B > A1 A2 A3 A1 A3 A2

  17. CES Solver Architecture 2. Continuous constraint propagation and consistency check. 4. Discrete constraint propagation and consistency check. 2.a. Propagation of additional constraints to discrete engine 4.a. Propagation of additional constraints to continuous engine 2.b. Propagation of additional constraints to continuous engine, repeat 2.a, 2.b as needed/desired 4.b. Propagation of additional constraints to discrete engine, repeat 4.a, 4.b as needed/desired 1. Continuous constraints added as result of discrete decisions Discrete Constraint Engine ContinuousConstraint Engine 3. Discrete constraints added as result of continuous decisions

  18. Data structures: Discrete Model • Discrete variable • Legal values • Current constraints on assignable values • Cross-domain constraints: constraints to be added to continuous domain, depending on value assigned to this variable. • Discrete Constraint • n-ary relation on discrete-valued variables (legal combinations of values) • Constraint type: REQUIREMENT, DECISION, PROPAGATION EFFECT • Culprit Identification bookkeeping (Decision variable(s) responsible for this constraint being added). • Propagation method(s)

  19. Data structures: continuous model • Continuous variable • Current constraints • Continuous Constraint • Mathematical relation (=, <=, etc.) on several variables • Constraint type: REQUIREMENT, DECISION, PROPAGATION EFFECT • Culprit Identification bookkeeping (Decision variable(s) responsible for this constraint being added). • Propagation method(s)

  20. Constraint Envelope Scheduling Advantages: • No premature commitment • Flexible search strategies • Natural representation for decisions (“do it before lunch”) • Natural interleaving of search and propagation Disadvantages: • (Somewhat) cumbersome to implement. • Intermediate results are less intuitive to users (more difficult to present, anyway). • Efficiency concerns (a matter of degree…)

  21. CES vs. Timeline Scheduling • Activity times and resource usage are precisely specified, if activities appear on the schedule at all, vs. • Activity extent and resource usage may be incrementally constrained. The first assumption makes propagation difficult. The second assumption can get you into trouble in a number of ways: • excessive (and potentially irrelevant) bookkeeping • decisions made in a context that is then invalidated (Andy Baker’s “Hazards of Fancy Backtracking”).

  22. More Tradeoffs • Constraints are checked when activities are added to the schedule, vs. • Constraints are accumulated as scheduling progresses The first assumption makes incremental rescheduling difficult, throws away information about how the schedule got to its current state. Under the second assumption, you can accumulate a lot of constraints. Most of them may be subsumed by other constraints, but employing any kind of backtracking search will require that they are maintained in some form.

  23. Implementing a Generic Scheduling Core • Generic CSP support • variables and constraints • propagation • search • Continuous model for temporal constraints • Activities • Resources • Resource requirements

  24. Activities • Interval • temporal constraints, in the Interval Constraint Engine (ICE) • Resource requirements • Hierarchical activities • Container activities

  25. Interval Constraint Engine (ICE) A graph of time points, with labeled edges between them. • Possible temporal relationships: • Release time and deadline (relative to a clock) • Min/max duration • Precedence (minimum separation) • Latency (maximum separation) • Ordering decisions and start/end time assignments are added as constraints. • ICE maintains global consistency, reports infeasible (negative-weight) cycles.

  26. Resources • Unary Resource (manufacturing cell) • Metric Resource (power) • Capacity resource (energy) • State resource (material service) • state-using activities • state-changing activities

  27. Upper and Lower Bounds on Resource Use

  28. Generalized Bounds Bounds based on temporal constraints are only part of the problem: • Activities on/off the schedule. • Resource choices unconstrained or partially constrained. • Imprecise usage by individual activities

  29. Resource Requirements • Match: • by name • by type • by attribute • by procedural attachment • Effect • metric resource is busy • capacity resource is consumed (or produced) at a constant rate. • state resource is changed (or not).

  30. Effective Domain Modeling is Hard Example: large transport pipelines • Pipeline volume must be modelled explicitly (not like shipping and receiving). • Currently model is to have slugs of material with associated volume, etc., plus a position in the pipeline. • Material movement in the pipeline requires both input and output flows. • Rate is independent of individual movements

  31. Boeing 777 AIMS Scheduling

  32. SAFEbus Scheduler for Boeing 777 AIMS • AIMS requirements represent one of the largest, most comprehensive sets of constraints ever successfully scheduled: • 29,000 items are scheduled, subject to 97,000 complex metric constraints specified by AIMS applications developers • More than 230,000 decisions, each with between 4 and 4,000 possible choices were made in scheduling • This corresponds to finding a solution with a directed search of 107 elements in a state space of 10140000.

  33. Processor Schedule

  34. Data Schedule

  35. How We Solved It Ginsberg’s “Dynamic Backtracking” algorithm • Published in 1994 • We fixed a bug • We extended it Moral: sometimes the path from research to product takes months, not years.

  36. “Current state of the art” Issues: 60% of manufacturing facilities have no automated scheduling capability. 30% of airline flights don’t go as planned, on a good day. 10% reduction in waiting time for trucking industry would allow 6% of all trucks to be taken off the highways altogether. Most current scheduling done by unsophisticated means - spreadsheets or manually. Gaps to be closed by an Integrated Solution: 1) Schedule awareness and editing, then… 2) Autogeneration of feasible schedules, then… 3) Autogeneration of optimal schedules. Scheduling: The Current State of the Art Most manufacturing enterprises do not have an effective solution to address Step 1!

  37. Scheduling Systems are 95% Boring Planning Tool Plan Ext Sched Marketing Archive Archived Schedule Config Published Schedule Scheduler PRM Plant Model Publish Plant Historian Plant Working Schedule Data Externally Accessible Activity Hist MES External Data Interchange Scheduling UI External Data Interchange

  38. Areas for Further (Interesting!) Work • Optimal Scheduling (for the right definition of “optimal”) • Incorporating probabilities • Robust schedules vs. responsive schedulers • Moving up into production planning • Better integration with action selection and states

More Related