1. Thesis Defense
 Investigation of Agent-Based Approaches to Enhance Container Terminal Operations by
Omor Sharif
Presented in Partial Fulfillment of the Requirements
For the Degree of Master of Science in
Civil Engineering
2011

2. What is a Container Terminal (CT)?
An interface between ocean and land
Ships are loaded and unloaded
Containers are temporarily stored
Manage handling of Containers etc

3. Flow of Containers and Decision Problems
Yard Operations - Storage Space Assignment
Berth Allocation
Quay Crane Scheduling
Yard Operations – Yard Crane Scheduling
Transport of Containers to Storage Area and Vice Versa
Delivery and Receipt Operations (Gate Operations)

4. Research Topics Two Research Studies Yard Crane Scheduling Problem
Truck Queuing at Terminal Gates
1. Sharif, O., Huynh H. (2011) “Yard crane scheduling at seaport container terminals: A comparative study of centralized and decentralized approaches”. Paper to be submitted for publication.
2. Sharif, O., Huynh, H., Vidal, J. (2011) “Application of El Farol model for managing marine terminal gate congestion”. Submitted to Journal of Research in Transportation Economics.

5. Journal Article I
 Yard Crane Scheduling at Seaport Container Terminals: A Comparative Study of Centralized and Decentralized Approaches  by
Omor Sharif and Nathan Huynh
University of South Carolina
Paper to be submitted for publication

6. Outline What is Yard Crane Scheduling Problem?
Review of Centralized Solution
Review of Decentralized Solution
Design of Experiments and Results
Comparative Performance between the two approaches
Conclusion/Future Directions

7. Yard Crane Scheduling Problem
Objective: Determining best sequence of trucks to serve by each yard crane.
Challenges:
There are fluctuations in truck arrival
Job locations are distributed throughout the yard zone
Good decisions are difficult to conceive manually

8. Yard Crane Scheduling (YCS) Problem
Motivation
Operational improvement of container terminal
Reducing drayage trucks turn time
Efficient allocation of scarce resources
Environmental Concerns

9. YCS Problem Solution Solution to YCS Problem Centralized Approaches
-OR Optimization
- IP
- MIP
Decentralized Approaches
- Agent-based Modeling

10. Research Questions Comparative Study between the two approaches
Contrasting assumptions?
Strengths and weaknesses?
Relative performances?
Suitability for implementation?

11. Centralized Approach Based on the work of Ng (2005)
IP was developed for optimal crane scheduling
Considers multiple yard cranes and known arrival times
Excessive computational time required to solve IP
Dynamic programming based heuristic is proposed

12. Centralized Approach First Phase (Find Best Partition)
How the Heuristic solves YCS?
Heuristic has TWO phases
First Phase (Find Best Partition)
Partitioning of the Yard Zone
Several smaller groups equal to number of YCs
Job handling follows greedy heuristic
Output is best partition with least total waiting

13. Centralized Approach Second Phase (Job Reassignment)
How the Heuristic solves YCS?
Heuristic has TWO phases
Second Phase (Job Reassignment)
Job reassignment between adjacent YCs
Interference check required
Algorithm considers two cranes at some time
Output is the minimum total waiting found by heuristic

14. Centralized Approach A Sample Heuristic Solution First Phase Solution
Second Phase
Solution
Path of the
Cranes

15. Decentralized Approach
Distributed perspective in recent years
Based on the work of Huynh and Vidal (2010)
Agent based approach
Each YC is an agent seeking to maximize utility
Decisions are based on the valuation of utility function
Utility functions are designed to minimize waiting time

16. Decentralized Approach
Utility Functions
Distance Based Utility
Time Based Utility
D = Distance to Truck
T = Truck Wait Time
p1 and p2 = Penalty Values (discouraging penalties)
Xinterference, Xproximity, Xturn and Xheading are binary variables

17. Decentralized Approach
Simulation model, coded in Netlogo
Netlogo: A multi-agent programmable Environment

18. Key Differences Centralized approach Decentralized approach
Optimization strategy
Global optimization.
Agent based local optimization.
Work flow
Optimal schedule.
Individual decisions.
Arrival information
Assumes complete information.
No assumption.
Truck sequencing
Greedy approach
Cranes’ utility functions.
Implement-ation
Dynamic heuristics.
Agent-based simulation.

19. Experimental Design A large set of YCS problems were solved
Experiment Set 1: Impact of Number of Yard Cranes
Number of YCs ⟶ 2 to 4
Experiment Set 2: Impact of Truck Arrival Rate
Arrival Rate ⟶ 5, 10 and 15
Experiment Set 3: Impact of Yard Size
Number of Yard blocks ⟶ 1 to 3
Experiment Set 4: Impact of Truck Volume
Number of Jobs ⟶ 20, 50 and 80
Job location distribution ⟶ Random Uniform Distribution
Job arrival distribution ⟶ Poisson Distribution

20. Comparative Performance between the two approaches
Optimality - Minimize the truck waiting time
Centralized Approach
Heuristic produces near-optimal schedule
On average 7.3% above the lower bound
Decentralized Approach
No advance schedule for the agents
On average 16.5% above the heuristic solution

21. Comparative Performance between the two approaches
Optimality - Minimize the truck waiting time
Fig: Mean Index for different truck arrival rates

22. Comparative Performance between the two approaches
Optimality - Minimize the truck waiting time
Fig: Mean Index for different yard sizes
Fig: Mean Index for different job volumes

23. Comparative Performance between the two approaches
Scalability and computational efficiency
Centralized Approach
Highly sensitive to the size and complexity
Requires performing the computation in advance
Decentralized Approach
No computation time required in advance
Disaggregated, handle large and complex problems

24. Comparative Performance between the two approaches
Adaptability
Centralized Approach
Assumes complete information on supply and demand
Requires rescheduling to adapt with changes
Decentralized Approach
No assumptions on the arrival-time of trucks
Monitor changes continuously, adapt rapidly

25. Concluding Remarks/ Future Work
Two approaches have complimentary solution properties
Hybrid approaches may offer better results
Proposed Hybrid Approach I
Local optimization models for cranes
Coordination for best partition within yard zone
Proposed Hybrid Approach II
Solve global optimization periodically
Switch to adaptive agent-based model when necessary

26. Journal Article II
 Application of El Farol Model for Managing Marine Terminal Gate Congestion  by
Omor Sharif , Nathan Huynh and Jose Vidal
University of South Carolina
Submitted to Journal of Research in Transportation Economics

27. Outline Gate Congestion problem at CT
Proposed Model and Implementation
Design of Experiments and Results
Concluding Remarks

28. Congestion Problem at Terminal Gates
Documentation processing, inspection, security checks etc
Long waiting time due large number of idling trucks
Impact turn around time of drayage trucks
Environmental concern due to significant emission

29. Solution to the Gate Congestion Problem
Attempted
Solution
Appointment Systems/ Reservation Systems with Time Windows
Real Time Gate Congestion Information Using Webcams

30. Proposed Agent-based Model

31. Proposed Agent-based Model (Contd.)
N ≡ Set of Depots (n ∈ N)
T ≡ Set of Trucks (t ∈ T)
L ≡ Tolerance (Max allowed waiting time)
E (W) ≡ Expected wait
SEND? (n, t) ≡ 1 if E (W) ≤ L
0 otherwise
Total time before entry into port = T (n, P) + Q(t) + S(t)
Wait at gate, W(t) = Q(t) + S(t)
I ≡ Discretization interval
Average waiting at xth interval,
Historyx = { }

32. Proposed Agent-based Model (Contd.)
Parameters related to ‘Predictors’
S = [s1, s2 ,s3 ,..., sz] ≡ Predictor space containing z predictors
k ≡ Number of predictors chosen from S
my-predictors-list(n) ≡ Predictor set of depot agent n
my-predictors-scores(n) ≡ Rank of predictors of depot agent n
my-predictors-estimates(n) ≡ for each predictor
sactive−predictor(n) ≡ Best performing predictor for depot agent n
Updating of scores
Original Precision Approach:
  is a number strictly between zero and one

33. Proposed Agent-based Model (Contd.)
Pseudo Code of the Program – Part of the Main Loop

34. Model Implementation
Simulation model, coded in Netlogo

35. Experimental Design Parameter Value Unit Number of Depots 10 Nos
Dispatch rate (θ) 12
trucks/depot/hr
Mean transaction time (μ)
3, 4, 5, 6, 7 and 8
minutes
Tolerance (L)
15, 20, 25 and 30
Total predictors
200
Predictors per depot (k)
Nos / depot
Update interval (I)
5,10 and 15
Maximum memory (m) 20
intervals
Predictor scoring policy
Original precision
n/a
Alpha
0.5

36. Results (Mean wait and Total completion)
Fig: Impact of tolerance on mean wait time of trucks
Fig: Impact of tolerance on total completion time.

37. Results (Mean wait time history)
Fig: Mean wait time of trucks (I =15 minutes, L = 15 minutes)
Fig: Mean wait time of trucks (I =10 minutes, L = 10 minutes)

38. Results (Base Case Comparison)
43% and 63% lower mean wait time for I = 5 and 10 mins
22% and 40% lower maximum wait time for I = 5 and 10 mins
18% and 40% higher completion time for I = 5 and 10 mins

39. Concluding Remarks Proposed model provides steady truck arrival
Adopt higher ‘I’ for distributing demand
Good amount of emission reduction over ‘do-nothing’
First study of its kind
Additional studies are required to understand complexity
More sophisticated learning models

40. Thank You Questions ?