Title: Vehicle Routing
1Vehicle Routing Job Shop Scheduling Whats
the Difference?
ICAPS03, June 13, 2003
J. Christopher Beck, Patrick Prosser, Evgeny
Selensky
Dept. of Computing Science University of Glasgow
pat,evgeny_at_dcs.gla.ac.uk
Cork Constraint Computation Centre University
College Cork c.beck_at_4c.ucc.ie
2Old Solutions for New Problems
- We have strong techniques to solve hard problems
- Use them!
- use existing problem models and solution
techniques to solve a new problem - Common approach in research and in practice
- SAT, IP, CP, etc
- If you have a hammer,
3A Nice Idea, But
- New problems dont fit exactly the old models
- New problems look strange
- Scheduling with 0 duration activities
- Routing with 0 travel time
- How will solution techniques work?
- is the problem really a nail?
4Get the Picture?
Existing Problem Models
5This Paper
- Basic Question
- How does existing solution technology cope with
changed characteristics? - Basic Approach
- Create problems between JSP VRP
- Compare the relative performance of routing and
scheduling solution techniques - What problem characteristics are important to the
solution techniques? - More in talk than in the paper
6Vehicle Routing Problem
- Make a set of deliveries (visits) with a set of
vehicles - Vehicles have limited capacity
- Visits have time windows
- Minimize total distance traveled
T1
T3
T2
7Job Shop Scheduling Problem (JSP)
R1
R0
R2
R1
R2
R0
R0
R1
R2
8Off-the-Shelf Solution Technology
- VRP ILOG Dispatcher
- First Solution Savings Heuristic
- Improvement Guided Local Search
- JSP ILOG Scheduler
- Constructive CP tree-search
- Slack-based heuristics
- Strong constraint propagation
- Edge-finding, precedence graph
9Evaluating the Technology
- Cx cost of solution found by technology x with
fixed time limit (10 minutes) - ? gt 1 routing technology is better
- ? lt 1 scheduling technology is better
10JSP ? VRP Transformation
- Beck et al. 2002
- We can transform JSPs to VRPs and vice versa
- Scheduling technology is poor on reformulated
VRPs - Routing technology is poor on reformulated JSPs
- Cant find first solutions due to precedence
constraints!
11Base Case Pure Problems
12Characteristics
- What are the problem characteristics that lead to
this difference? - Ideas
- Alternative resources
- Optimization criteria
- Precedence constraints
- (3 more not really discussed here)
13From VRP
??
VRP
14From JSP
??
JSP
15Alternative Resources
- VRP many (e.g., 25)
- JSP few (1, 4, 8)
- Savings cant solve 70 of problems with 2
alternatives - Only problems solved by both are included
16Alternative Resources VRP
17Alternative Resources JSP
18Optimization Criteria
- VRP total travel
- JSP makespan
19Optimisation Criteria VRP
20Optimisation Criteria JSP
21Precedence Constraints
- VRP none
- JSP paths of totally ordered activities
- Savings cant find first solution
- Start with scheduling solution
22Precedence Constraints VRP
23Precedence Constraints JSP
24Experimental Summary
JSP
VRP
? Alt Res ?
? Alt Res ?
? Precedence Cts ?
? Precedence Cts ?
Opt. Makespan
Opt. Total Travel
? scheduling performance
? routing performance
25Other Characteristics
- Smaller impact
- Temporal Slack
- ? slack ? scheduling performance
- Vehicle Resource Capacity
- Like alternative resources
- Activity duration to transition time ratio
- VRP ? ratio ? routing performance
- JSP ? ratio ? scheduling performance
26Conclusions
- Try scheduling technology on VRP with
- makespan minimization (strong propagation?)
- complex temporal constraints
- Try routing technology on JSP with
- total time minimization (weak propagation?)
- few temporal constraints (open shop?)
27Conclusions
- Even isolated changes in problem characteristics
change the best choice of off-the-shelf problem
model - Understanding this is important to extending the
scope of optimisation techniques to - new problems
- new people
28Alternative Resources