21
Cooperative Co-evolutionary Optimisation on Work Package Scheduling and Staff Assignments Jian Ren, Mark Harman CREST Centre, SSE Group, UCL 23 Feb, 2011, 11 th COW

Jian Ren, Mark Harman CREST Centre, SSE Group, UCL 23 Feb, 2011, 11 th COW

Embed Size (px)

Citation preview

Page 1: Jian Ren, Mark Harman CREST Centre, SSE Group, UCL 23 Feb, 2011, 11 th COW

Cooperative Co-evolutionary Optimisation on Work Package Scheduling and Staff Assignments

Jian Ren, Mark Harman

CREST Centre, SSE Group, UCL23 Feb, 2011, 11th COW

Page 2: Jian Ren, Mark Harman CREST Centre, SSE Group, UCL 23 Feb, 2011, 11 th COW

Outline

Research MotivationsSolutions Representations: WPO and StaffingFitness Evaluation: Simulation of ExecutionCooperative Coevolution ProcessEmpirical Results and Issues

Page 3: Jian Ren, Mark Harman CREST Centre, SSE Group, UCL 23 Feb, 2011, 11 th COW

Motivations

Staff Assignments (Team Construction)How to assign staff into project teams?

Work Packages Scheduling (WP Ordering)How to put the WPs in a good order to execute?

ObjectiveFind earlier completion time of a project by optimising TC and WPO simultaneously

Page 4: Jian Ren, Mark Harman CREST Centre, SSE Group, UCL 23 Feb, 2011, 11 th COW

Solution Representation of TC How to assign N staff into M teams?

Staff1

Staff 2

Staff3

Staff4

Staff5

… Staff29

Staff 30

To Team

3 4 2 3 5 … 3 2

N staff

Species #1: assignment of staff to teams

Page 5: Jian Ren, Mark Harman CREST Centre, SSE Group, UCL 23 Feb, 2011, 11 th COW

Cooperative Coevolutionary Optimisation on Work Package Scheduling and Staff Assignments Jian Ren & Mark Harman

5

Genetic Operator on TC

3 4 2 3 5 … 3 2Parent_1

Page 6: Jian Ren, Mark Harman CREST Centre, SSE Group, UCL 23 Feb, 2011, 11 th COW

Cooperative Coevolutionary Optimisation on Work Package Scheduling and Staff Assignments Jian Ren & Mark Harman

6

Genetic Operator on TC

3 4 2 3 5 … 3 2

4 2 5 2 4 … 2 1

4 2 5 2 5 … 3 2

3 4 2 3 4 … 2 1

Parent_1

Parent_2

Child_1

Child_2

Single Point Crossover

Page 7: Jian Ren, Mark Harman CREST Centre, SSE Group, UCL 23 Feb, 2011, 11 th COW

Solution Representation of WPOHow to put the WPs in a perfect or near perfect order?

Perfect ordering = All teams are busy all the time and finish their last WP at the same time

Solution:

Attributes of WP ID Efforts: total efforts required by executing the WP Dependency: ID of its predecessor

WP1 WP4 WP6 … WP56 WP60

Species #2: orderings of the WPs

Page 8: Jian Ren, Mark Harman CREST Centre, SSE Group, UCL 23 Feb, 2011, 11 th COW

Cooperative Coevolutionary Optimisation on Work Package Scheduling and Staff Assignments Jian Ren & Mark Harman

9

Genetic Operator on WPO

WP1 WP4 WP6 … WP56 WP60Parent_1

Page 9: Jian Ren, Mark Harman CREST Centre, SSE Group, UCL 23 Feb, 2011, 11 th COW

Cooperative Coevolutionary Optimisation on Work Package Scheduling and Staff Assignments Jian Ren & Mark Harman

10

Genetic Operator on WPO

WP1 WP4 WP6 … WP56 WP60Parent_1

WP2 WP3 WP1 … WP60 WP59Parent_2

WP1 WP4 WP6 … WP60 WP59Child_1

WP2 WP3 WP1 … WP56 WP60Child_2

B: the rest of the other parent after WPs in A are deleted

A: copied from one parent

Page 10: Jian Ren, Mark Harman CREST Centre, SSE Group, UCL 23 Feb, 2011, 11 th COW

Fitness Evaluation

Objective Earlier Overall Completion Time = Higher Fitness

WP1 WP4 … WP60

Efforts

8 4 12

Processing Simulator The `Best` available team picks the next WP before other available teams. First Come First Served One WP blocks the queue of waiting WPs until its predecessors are all finished.

Staff1

Staff2

Staff3

… Staffn-1

Staff n

To Team

3 4 2 … 3 2

Page 11: Jian Ren, Mark Harman CREST Centre, SSE Group, UCL 23 Feb, 2011, 11 th COW

Processing Simulator

Fitness Evaluation

Objective Earlier Overall Completion Time = Higher Fitness

T1 T2 … T5

Capacity 6 staff 8 staff … 7 staff

Date (it becomes available

)

0 0 … 0

WP1 WP4 … WP60

Efforts

8 4 12

StartTime

FinishTime

To Team

Page 12: Jian Ren, Mark Harman CREST Centre, SSE Group, UCL 23 Feb, 2011, 11 th COW

Processing Simulator

Fitness Evaluation

Objective Earlier Overall Completion Time = Higher Fitness

T1 T2 … T5

Capacity 6 staff 8 staff … 7 staff

Date (it becomes available

)

0 1st Day

… 0

WP1 WP4 … WP60

Efforts

8 4 12

StartTime

0th Day

FinishTime

1st Day

To Team

T2

Page 13: Jian Ren, Mark Harman CREST Centre, SSE Group, UCL 23 Feb, 2011, 11 th COW

Processing Simulator

Fitness Evaluation

Objective Earlier Overall Completion Time = Higher Fitness

T1 T2 … T5

Capacity 6 staff 8 staff … 7 staff

Date (it becomes available

)

22nd Day

21st Day

… 21st Day

WP1 WP4 … WP60

Efforts

8 4 12

StartTime

0th Day

0th Day

20th Day

FinishTime

1st Day

0.5th Day

22nd Day

To Team

T2 T3 T1

The latest finishing time is the overall completion time of a simulation.

Page 14: Jian Ren, Mark Harman CREST Centre, SSE Group, UCL 23 Feb, 2011, 11 th COW

Cooperative Coevolutionary ProcessInitialisation for both species //(S1: TC, S2: WPO)FOR (E = 1 : External_Generations)

FOR (I = 1 : Internal_Generations)internal evolutionary progress for

Species #1end FORFOR (I = 1 : Internal_Generations)

internal evolutionary progress for Species #2

end FORend FOR

Page 15: Jian Ren, Mark Harman CREST Centre, SSE Group, UCL 23 Feb, 2011, 11 th COW

Cooperative Coevolutionary Optimisation on Work Package Scheduling and Staff Assignments Jian Ren & Mark Harman

16

Cooperative Co-Evolution Algorithm

Internal evolutionary

process for TC

Internal evolutionary process for

WPO

Evolved TC Evolved WPO

Page 16: Jian Ren, Mark Harman CREST Centre, SSE Group, UCL 23 Feb, 2011, 11 th COW

18

Coevolutionary Process on Boxplot

Internal Generation Num = 10External Generation Num = 10X: Generation Num (Time)Y: Project Completion Time

Page 17: Jian Ren, Mark Harman CREST Centre, SSE Group, UCL 23 Feb, 2011, 11 th COW

Given the same number of evaluation efforts, CCEA finds better solutions than single population EA.

Optimisation on WPO contributes more benefit than on TC.

Preliminary Empirical Results

Page 18: Jian Ren, Mark Harman CREST Centre, SSE Group, UCL 23 Feb, 2011, 11 th COW

Cooperative Coevolutionary Optimisation on Work Package Scheduling and Staff Assignments Jian Ren & Mark Harman

20

Configurations for CCEA

Internal Generation Num

External Generation Num

I 1 100

II 10 10

III 100 1

CCEA finds better solutions than single-species EA

* Configuration III can be considered as single-species EA, as the FOR loop has only been executed once. No interactions between two population have chance to happy before the end of the FOR loop. It is an extreme case of CCEA.

3 sets of configurations are applied to 4 real-world projects.Comparing the earliest finishing time found by I, II and III:

Configurations’ Achievement

#First Place

#Second Place

#Third Place

I 6 1 1

II 2 6 0

III 0 1 7

Page 19: Jian Ren, Mark Harman CREST Centre, SSE Group, UCL 23 Feb, 2011, 11 th COW

21

Optimisation on WPO contributes more benefit than on TCInternal Generation Num =

100External Generation Num = 1X: Generation Num (Time)Y: Project Completion Time

Page 20: Jian Ren, Mark Harman CREST Centre, SSE Group, UCL 23 Feb, 2011, 11 th COW

Cooperative Coevolutionary Optimisation on Work Package Scheduling and Staff Assignments Jian Ren & Mark Harman

22

Given the same number of evaluation efforts, CCEA finds better solutions than single population EA.

Optimisation on WPO contributes more benefit than on TC.

Different configurations of the CCEA produce results with different characteristics.

Preliminary Empirical Results

Page 21: Jian Ren, Mark Harman CREST Centre, SSE Group, UCL 23 Feb, 2011, 11 th COW

Cooperative Coevolutionary Optimisation on Work Package Scheduling and Staff Assignments Jian Ren & Mark Harman

23

Problem ModelStaffing

skills, changing teams during the project, absence, etcWP

changed during the project, mis-estimate efforts

The configurations of CCEA is complex.Diversity of the species of Team Constructions,

Processing Simulator (execution of WPs)Job-shop scheduling permutation, etc

Comparison and ValidationSingle-species EA (Statistics Analysis)

Verification by More Real-world Data fromSoftware Project Management, Cloud Computing

Issues & Future Work