Transcript policy evolution within an organization

Policy Evolution within
an Organization
James H. Hines
Sloan School of Business, Massachusetts
Institute of Technology
Jody Lee House
Department of Electrical and Computer
Engineering, Oregon Graduate Institute
Funded in part by NSF IOC Award#SES-9975942
The Problem
System-wide company improvement is
difficult because companies are too
complex to “solve.”
How can we improve organizations in the
face of ignorance?
A solution?


Biological evolution has produced
excellent organizations.
Can we identify analogs of natural
evolution that will help human
organizations to likewise excel?
Gene:Organism::Policy:Organization
Policy:
Implicit or explicit
 Examples






Pricing
Hiring
Capacity Expansion
Flywheel sales
Synonyms


Decision rule
Rule of thumb
A policy produces A gene produces
• A stream of decisions
• A stream of proteins
• Activity in the firm
• Activity in the cell
• Changing the policies,
changes the
organization
• Changing the genes
changes the organism
Where are “evolutionary
packets” stored?
• Genes are stored on chromosomes in
cells
• Policies are stored
– In written manuals?
– In committees?
– On computers?
– In brains of people
Processes
Genes



vs
Mutation
Recombination
Natural selection
and the sex drive

survival of the
fittest
Policies



Innovation
Inter-personal learning
Pointing and pushing
mechanisms

learning from the fittest
Pointing And Pushing
Mechanisms



Point to successful people
Push others to learn from them
Examples




Promotion and hierarchy
Pay scales
The best and latest computers
In house training?
A brief look at sex
Grandpa’s
strand
Papa
Grandma’s
strand Mama
cell
chromosome
recombination
sperm
egg
you
fertilization
recombination
Recombination is key
• Combine parts of fit organisms to create fitter
organism
• Example: 4-digit number, A > B
fitter
8,765
7,999
8,999
Learning is Similar to Biological
Recombination
Fred
Phyllis
brain
Time 1
policy
Phyllis teaching
Fred learning
Time 2
Why learning is difficult to call
to mind
• The donor’s idea is well integrated
• The rest of the donor’s idea is difficult to
recognize as an idea
Overview
Step 3: Promote
Managers
Step 2: Evaluate
performance of the
system dynamics
models
Step 1: Run system
dynamics simulation
models, using policies of
the managers
Step 4: If using
teams: Mix
managers and
reform teams
Step 5: Managers
learn
Step 6: Managers
innovate
Step 1: Run
SD models
Decides
Correct code
Programmers
Decides
Productivity
Writing code
correctly
Decides
Code to write
Correct code
Programmers
Writing code
correctly
Correct code
Programmers
Productivity
quality
WritingCode
Productivity
Creating
bugs
Code to write
qualityUndiscovered bugs
WritingCode
Writing
code
correctly
Code to write
quality
WritingCode
Creating
DiscoveringBugs
bugs
Undiscovered bugs
DiscoveringBugs
BugDiscoveryTime
DiscoveringBugs
Creating
bugs
Undiscovered bugs
BugDiscoveryTime
BugDiscoveryTime
Step 1 The Project Model Detail
DesiredPeople
HireFire
Rate
Remaining
Time
<Time>
timeTo
Change
WF
People
Productivity
WorkToDo
DueDate
Correctly
Doing
Doing
Correctly
Done
Normal
Quality
Quality
UndiscoveredBugs
Anticipated
TimeTo
IncorrectlyDoing
Complete
TimeTo Anticipated
BugDetecting
DueDate Anticipated
Change
TimeToDetect
Production
Rate
Schedule
Bugs
<People>
<Time>
<Productivity>
Step 2: Evaluating Performance



Fitness function can be based on any
variables in the model
Variables can be combined using any
functional form
In the following we use two simple fitness
functions


Time to ship (LastPossible – Actual)
Number of bugs (LinesOfCode – BuggyLines)
Step 3: Promoting managers
1. Rank individuals based on relative
performance
2. Promote according to rank.
Positionnew  Positionold * PromotionFactor
The promotion algorithm requires specifying the “promotion base”. A
promotion base of 2 means
•The highest performing manager’s new position is 2 * theOld
•The lowest performing manager’s new position is (1/2) * theOld
•Everyone else’s promotion is evenly spread out between 2 and 1/2
Step 3 Promotion Algorithm Detail
Positionnew  Positionold * PromotionFactor
PromotionFactor  baseValue
K
2 * (rank  1)
K
1
populationSize  1
Team-based promotion
Positioni ,new  Positioni ,old * TeamAPromotionFactor
Step 4: If using teams mix them up



Randomly?
Spread out the best?
Concentrate the best?
Step 5: Learn
a) Select a teacher by
roulette
b) Learn from the teacher
by recombination
Pos= Manager5
0.5
Position=1.41
Manager3
Position=2
Manager1
Pos=1
Pos=0.71
Manager2
Step 5 Learn  p(learn)
OLD
Learner’s Policy
10 or 0010 10
Teacher’s Policy
32 or 1000 00
Randomly choose a
crossover Point, say 2
Randomly choose which part the learner will obtain and which he
will retain
0010__ + ____00
0010 00 = 8
OR
____10 + 1000__
1000 10 = 34

NEW
Learner’s Policy
34 or 100 10
Teacher’s Policy
32 or 1000 00
Step 6: Innovate  p(innovate)
111111
Before
Flip !
110111
After
Learning, no pushing/pointing:
Learning Drift
Optimal value = 8
Learning, no pushing/pointing
Random Consensus
Policy
15
10
5
0
0
5
10
Generation
15
20
Learning with Pointing/Pushing
Individuals
16
14
Policy
12
10
8
6
4
2
0
0
5
10
Generation
15
20
Next steps

Measurement through knowledge
elicitation with partner companies



Who learns from who and why?
How are implicit policies a function of
organizational structure?
Integrated simulation