Академический Документы
Профессиональный Документы
Культура Документы
Stephen F. Smith
The Robotics Institute
Carnegie Mellon University
Pittsburgh PA 15213
sfs@cs.cmu.edu
4122688811
Carnegie Mellon
Outline
• What is Scheduling?
• Current State of the Art: ConstraintBased
Scheduling Models
• Is Scheduling a Solved Problem?
Carnegie Mellon
What is Scheduling?
Allocation of resources to activities over time so
that input demands are met in a timely and
costeffective manner
Most typically, this involves determining a set
of activity start and end times, together with
resource assignments, which
• satisfy all temporal constraints on activity
execution (following from process
considerations)
• satisfy resource capacity constraints, and
• optimize some set of performance objectives
to the extent possible
Carnegie Mellon
A Basic Scheduling Problem
rel1 dd1
R1 R2
i j st(i) + p(i) < st(j), where p(i)
is the processing time of op i
op21 op22 i R j
st(i) + p(i) < st(j) st(j) + p(j) < st(i)
rel 2 dd2
rel < st(i), for each op i of job j
j
dd > st(i) + p(i), for each op i of job j
j
Carnegie Mellon
A More Complex Scheduling
Problem
Origin
SeaPOE
AirPOD
Carnegie Mellon
Scheduling Research:
The Last 10 Years
• Major advances in techniques for solving
practical problems
• Constraint solving frameworks
• Incremental mathematical programming
models
• Metaheuristic search procedures
• Several significant success stories
• Commercial enterprises and tools
Carnegie Mellon
ConstraintBased Scheduling
Models
Components:
Properties:
• Modeling Generality/Expressiveness
• Incrementality
• Compositional
Carnegie Mellon
What is a CSP?
Given a triple {V,D,C}, where
• V = set of decision variables
• D = set of domains for variables in V
• C = set of constraints on the values of
variables in V
Find a consistent assignment of values to
all variables in V
Carnegie Mellon
A Basic CSP Procedure
1. [Consistency Enforcement] Propagate
constraints to establish the current set vd of
feasible values for each unassigned variable d
2. If vd = Ø for any variable d , backtrack
3. If no unassigned variables or no consistent
assignments for all variables, quit; Otherwise
4. [Variable Ordering] Select an unassigned
variable d to assign
5. [Value Ordering] Select a value from vd to
assign to d.
6. Go to step 1
Carnegie Mellon
Formulating Scheduling
Problems as CSPs
“Fixed times” model
• Find a consistent assignment of start
times to activities
• Variables are activity start times
Disjunctive graph model
• Post sufficient additional precedence
constraints between pairs of activities to
eliminate resource contention
• Variables are ordering decisions
Carnegie Mellon
A Simple Job Shop Scheduling
CSP
Variables: start times (stj,i) Domain: [0,12]
R1 R3
Job2 : O2,1 O2,2
[0,12] [0,12]
R2
Job3 : O3,1 O3,2 O3,3
[0,12] [0,12] [0,12]
Carnegie Mellon
Constraint Propagation
Deductive process of inferring additional
constraints from existing constraints as
decisions are made
Two roles:
• Early pruning of the search space by eliminating
infeasible assignments
• Detection of constraint conflicts
Carnegie Mellon
Some Constraint Propagation
Terminology
Kconsistency guarantees that any locally consistent
instantiation of (K1) variables is extensible to any Kth
variable
Example: 2consistency (“arcconsistency”)
Complexity: Enforcing Kconsistency is (in general)
exponential in K
• Forward Checking: partial arcconsistency only
involving constraints between an instantiated
variable and a noninstantiated one
Carnegie Mellon
Temporal Constraint Propagation
through Precedence Constraints
Assume dui,j = 3 for all Oi,j
• Before propagation:
O1,1 O1,2 O1,3
[0,12] [0,12] [0,12]
• Forward propagation
O1,1 O1,2 O1,3
[0,12] [3,12] [6,12]
• Backward propagation
Carnegie Mellon
Capacity Constraint Propagation
Observation: Enforcing consistency with respect
to capacity constraints is more difficult due to
the disjunctive nature of these constraints
Forward Checking:
O1,1 Scheduled to start at time 6
[6,6]
R1
Before propagation: [6,12]
O2,1 After propagation: [9,12]
Carnegie Mellon
Pruning Operation Ordering
Alternatives
Example: Erschler’s dominance conditions
Conclusion: Oi cannot precede Oj
In general: For any unordered pair of operations {Oi, Oj},
we have four possible cases:
1. LSTi < EFTj and LSTj ≥ EFTi: Oi is before Oj
2. LSTj < EFTi and LSTi ≥ EFTj : Oj is before Oi
3. LSTi < EFTj and LSTj < EFTi : inconsistency
Carnegie Mellon4. LSTi ≥ EFTj and LSTj ≥ EFTi: both options remain open
Edge Finding
• S a set of operations competing for resource R
• O an operation not in S also requiring R
((LFT(S) EST(S) < Dur(O) + Dur(S))
EST(O) ≥ EST(S) + Dur(S)
(LFT(S) EST(O) < Dur(O) + EST(O))
OP k
OP j
OP i
10 20 30
S = {OP ,OP }; O = OP Start Time OP ≥ 25
i j k k
Carnegie Mellon
More Complex Temporal
Constraints
“Simple Temporal Problem” (STP) [Dechter91]
• Edgeweighted graph of time points expressing
constraints of the form: a≤ tpj− tpi≤ b
• Assuming no disjunction, allows incorporation of
• Temporal relations:
•finishtostart <0, ∞> (precedence)
•starttofinish <t1,t2> (duration)
•Starttostart <0,0> (samestart)
•...
• Metric bounds: offsets from time origin
• Efficiently solved via allpairs shortest path
Carnegie Mellon
ConstraintPosting Scheduling
Models
• Conduct search in the space of ordering
decisions
• variables Ordering(i,j,R) for operations
i and j contending for resource R
• values i before j, j before i
• Constraint posting and propagation in
the underlying temporal constraint
network (time points and distances)
Carnegie Mellon
Search Heuristics
(Variable and Value Ordering)
• Slack/Temporal Flexibility
• Choose pair of activities with least sequencing
flexibility
• Post sequencing constraint that leaves the most
slack
• Resource Demand/Contention
• Identify bottleneck resource
• Schedule (or sequence) those activities
contributing most to demand
• Minimal critical sets
• Generalization to multicapacity resources
Carnegie Mellon
Search Control
• Backtrackingbased search
• LeastDiscrepancy Search
• Iterative Restarting with randomized
heuristics
• Local search Tabu, GAs, etc.
Carnegie Mellon
The Broader Picture
Constraint posting provides a framework for
integrating planning and scheduling
• contemporary temporal planners operate with
analogous representational assumptions
• E.g., IXTET, HSTS/RAX, COMIREM, …
• “ConstraintBased Interval Planning” [D. Smith 00]
Constraint posting is a relatively unexplored approach
to scheduling with several advantages
• more flexible solutions
• simple heuristics can yield high performance
solution techniques under a wide variety of problem
constraints
Carnegie Mellon
Technological Strengths
• Scalability
• Modeling flexibility
• Optimization
• Configurable
So, Is scheduling a solved problem?
Carnegie Mellon
What is Scheduling (Again)?
Classic view:
• Scheduling is a puzzle solving activity
• Given problem constraints and objective criterion,
figure out how to best tile the capacity over time
surface with operations
• Research agenda specify new puzzles and/or
provide new best solutions
i j
OP1,1 OP1,2 OP1,3
st(i) + p(i) ≤ st(j), where p(i)
rd1 is the processing time of op i
dd1 R1 OP2,1 OP1,2
R1 R2
i R j
rd(j) ≤ st(i) for each op i of job j
rd2 dd2
Minimize ∑ |c(j) dd(j)|
Carnegie Mellon
What’s Missing from the
Classical View of Scheduling
Practical problems can rarely be formulated as
static optimization tasks
• Ongoing iterative process
• Situated in a larger problemsolving context
• Dynamic, unpredictable environment
Carnegie Mellon
Managing Change
“Scheduling” is really an ongoing process of
responding to change
Manufacturing
Project Crisis Action
Management Planning
• Unpredictable, Dynamic • Predictable, Stable
Environment Environment
• Robust response • Optimized plans
Carnegie Mellon
Approaches to Managing Change
• Build schedules that retain flexibility
• Produce schedules that promote localized
recovery
• Incremental rescheduling techniques (e.g.,
that consider “continuity” as an objective
criteria)
• Selfscheduling control systems
Carnegie Mellon
Incremental Schedule Repair
Several competing approaches to maintaining
solution stability
• Minimally disruptive schedule revision (temporal
delay, resource area, etc.)
• Prioritybased change
• Regeneration with preference for same decisions
Little understanding of how these techniques
stack up against each other
Even less understanding of how to trade stability
concerns off against (re)optimization needs
Carnegie Mellon
DelayedCommitment
Scheduling Procedures
Identify a contention peak and post a leveling constraint
Activity 2 Activity 2
R1 Activity 1
R1 Activity 1
Advantages
• Retain flexibility implied by problem
constraints (time and capacity)
• Can establish conditions for guaranteed
executability
Carnegie Mellon Carnegie Mellon
Building Robust Schedules
Some open questions:
• Extended conditions for “Dispatchability”
• Robustness versus optimization
• Use of knowledge about domain
uncertainties
• Local search with robust representations
Carnegie Mellon
SelfScheduling Systems
• Distribute decisionmaking among
individual entities (machines, tools, parts,
operators; manufacturers, suppliers)
• Specify local behaviors and protocols for
interaction
• Robust, emergent global behavior
Carnegie Mellon
Morley’s GM Paint Shop System
Paint
Bid Booth Bid parameters:
1
same color as
Dispatcher Announcement last truck
(new truck) space in queue
empty queue
Paint
Bid Booth
“If bid for same color then award
else if empty booth then award 2
else if queue space then award”
Carnegie Mellon
Tradeoffs
Advantages:
• Complexity reduction
• Simple, configurable software systems
• Robust to component failures
• More stable computational load
Problems:
• No understanding of global optima (or how to
achieve global behavior that attends to specific
performance goals)
• Prediction only at aggregate level (can become
unstable)
Carnegie Mellon
Adaptive Systems:
“Routing Wasps” in the Factory
Machine
1 RWasp P(route|ST,ØT) = _________ S T
2
Agent1
ST2 + ØT2
Machine
2
.. RWasp
Agent2
B C
. A Jobs B
RWasp A
Machine AgentN C Stimulus:
N B SB
Response Thresholds:
ØA, ØB, ØC, ...
Carnegie Mellon
Updating Response Thresholds
ØT = ØT – ∆1 if next job is same type as current job
ØT = ØT + ∆2 if next job is a different type
ØT = ØT – ∆3 if the machine is currently idle
• Routing framework can be seen as an adaptive variant
of Morley’s bidding rule
• Experimental results showing significant performance
improvement
Carnegie Mellon
Some Open Issues in Multi
Agent Scheduling
• Selfscheduling approaches do not preclude the
use of advance schedules
• How to incorporate?
• Opportunistic optimization
• Cooperative, distributed scheduling is a fact of
life in many domains (geographic constraints,
autonomous business entities, etc.)
• How to negotiate and compromise?
• Can selfinterest be compatible with global
performance objectives?
Carnegie Mellon
Integrating Planning & Scheduling
“Planning & scheduling are rarely separable”
Waterfall Model
Planner
Plan Schedule
Scheduler
Mixed-Initiative Model
Plan
Planner
Scheduler
Schedule
Carnegie Mellon
Design Issues
• Integrated search space versus separable sub
spaces
• Single solver versus interacting solvers
• Resourcedriven versus strategydriven
• Loose coupling versus tight coupling
Carnegie Mellon
JFACC Planner/Scheduler
Plan Server
PLANS Constraint
HTN
SCHEDULES Based
Planner
ANNOTATIONS Scheduler
(SRI)
(CMU)
TRIGGERS
Technological Experimental
• Interleaved generation & • Simple, lowcost info. exchanges yield
repair of plans/schedules • Marked reduction in comp. time
• Distributed architecture to • Comparable plan/schedule quality
support remote collaboration • More complex models can improve
performance further
Carnegie Mellon SRI International
Some Challenges that Remain
• Scheduling models that incorporate richer
models of state
• Can integrated P & S problems really be
solved as one big optimization task?
• The limitations of SATstyle approaches
• How to achieve tighter interleaving of action
selection and resource allocation processes
• Managing change in this larger arena
Carnegie Mellon
Requirements Analysis
“Scheduling is really a process of getting the
constraints right”
Current tools designed around a “Specify and
Solve” model of user/system interaction
• Inefficient problem solving cycle
MixedInitiative solution models
• Incremental solution of relaxed problems
• Iterative adjustment of problem constraints,
preferences, priorities
Carnegie Mellon
Use of Relaxed Models to Identify
Resource Capacity Shortfalls
Carnegie Mellon
The AMC Barrel Allocator
Domain: DaytoDay Management of Airlift & Tanker
Assets at the USAF Air Mobility Command (AMC)
Technical Capabilities:
• Efficient generation of airlift and tanker schedules
• Incremental solution change to accommodate new
missions and changes in resource availability over
time
• Flexible control over degree of automation
• Selective, usercontrolled constraint relaxation and
option generation when constraints cannot be
satisfied
Carnegie Mellon
Parameterizable Search
Procedures
AssignMission:
C141, [t1,t2]
Configuration
Search Configurations
GenResources
Feasible <GenRequestedRes,GenIntervals,EvalMinCompletion >
305th 437th 60th 62nd Delay <GenRequestedRes, GenDelayInts, EvalMinTardiness>
AMW AMW AW AW
OverAllocate <GenRequestedRes, GenOverInts,
GenIntervals
EvalMinOverUsage>
I1,305 I2,305 ... I1,437 I2,437 ... I1,60 I1,60 ... I1,62 ... Bump – <GenRequestedRes, GenBumpInts, EvalMinDisruption>
AlternativeMDS < GenAlternRes, GenIntervals,
EvalMinCompletion>
...
EvalCriteria Composite Relaxations …
Feasible <GenResources, GenIntervals, EvalMinCompletion>
Carnegie Mellon
Generate
Relaxation
Options
Carnegie Mellon
MixedInitiative Scheduling
Challenges
• Management of user context across decision
cycles
• Explanation of scheduling decisions
• Why did you do this?
• Why didn’t you do that?
• Adjustable autonomy
Carnegie Mellon
Research Directions for the Next
10 Years
• Deeper integration of AI and OR techniques
• Robust schedules and scheduling
• Global coherence through local interaction
• Extension to largerscoped problemsolving
processes
• Rapid construction of high performance
scheduling services
Carnegie Mellon