GTES UTC 2014

Yves Caseau – 2014 – Serious Games as a Tool to Understand Complexity in Market Competition 1/42

Serious Games as a Tool to Understand

Complexity in Market Competition: An

Evolutionary Game Theory Simulation Platform

November 28th, 2014 – v0.3

UTC – Labex MS2T

Yves Caseau

National Academy of Technologies –

AXA


Outline

Part 1: Motivations – Making sense in a complex world

Is there a better tool

than Excel™ ?

Part 2: GTES (Game-Theoretical Evolutionary Simulation)

Part 3: Smart Grid Systemic Simulation Example

Part 4: Mass Market Telephony Simulation Examples


Complexity is everywhere in our companies

Complexity is everywhere

Multiple elements and multiple relations,

emergent behavior (ecosystems)

Feedback loops and delays

Uncertainty

Planning / Forecasting is still a major corporate activity

Budget, marketing, business plans, …

Excel™ is still the preferred tool

Complicated spreadsheets …

at best, a few scenarios and sensitivity analysis

Today’s business practices are suited to a complicated world,

not a complex world

Taking competitors & markets into account (adaptation)

Taking uncertainty into account

Enough of linear extrapolations !

Part

I :

Moti

vati

ons


From Strategic Planning to Serious Games

The solution is not better forecasting

With stochastic approaches towards uncertainty …

With multi-variable optimization …

That’s what experience and complexity theory say

We need to develop skills to better prepare for whatever the

future is bringing (situation potential)

Cf. the Art of military war games

Practice of multiple simulated situations develop

reactive skills (reflexes) and systemic understanding

Lessons from multiple strategic thinkers (Julien, Taleb, …)

Serious Games

Play against “smart” opponents

Experience feedback loops

Each scenario (game) is plausible, even if not likely

Part

I :

Moti

vati

ons


History of GTES Development

2000: UMTS Bid

2004 – 2006 : Distribution Channels Optimization

2006 – 2009 : Mobile Operator Competition Model

2009 – 2010 : Extension to Free

2010 – 2012 : Smart Grid Model

Reconcile three geographic visions of Smart Grids

Reconcile two corporate visions of Smart Grid

Part

I :

Moti

vati

ons

The « Utility »

view

The “Utility view” defines a smart grid as

adapting the power network to:

• local sources (as opposed to a one-way

distribution network),

• intermittent production sources (though

storage and favoring flexible production

units)

• using price incentives to “shave”

demand peaks.

The « Google »

view

The “Google view” defines a smart grid as:

• change from a tree structure to

a network structure (centralized to de-

centralized),

• the use of market forces to create

a dynamic and more efficient

equilibrium between supply and demand,

• the use of IT to provide information to

all actors, including end consumers.

The « Japanese »

view

The “Japanese view” is human-centered

instead of being techno-centered. The

goal is to change human behavior to

adapt to new challenges (lack of

resources, global warming, …). Smart

grids are the backbone of a multi-scale

architecture (smart home,

neighborhood, city, region, country)

where each level has its own resources

and autonomy.


Systemic Simulation of Smart Grid

« Regulator »

• Strategy: reduce CO2 emissions, preserveeconomic throughput (enough energy at« acceptable price »), keep a balancedbudget

• Tactical play: incentives to invest in greenenergy, CO2 Tax, storage requirement forintermittent sources of energy Open Questions

• What part does local

storage play?

• What CO2 price would

change significantly the

cost/benefits analysis?

• What is the systemic

benefit of local

management?

• What could be the large-

scale effect of dynamic

pricing on self-

optimization of customer

demand?

• Does Smart Grids provide

better resilience ?

• Is the relationship

between supplier and

operators a “coopetition”

or a competition ?

« Supplier »

• Strategy: maintain EBIDTA, reduce exposureto demand peaks, maintain market share

• Tactical play: Variable pricing (higher pricewhen demand & production costs are high),power plant investments

« Operator »

• Strategy: grow turnover, grow EBITDA,increase market share

• Tactical play: Storage utilization policy,Dynamic pricing, when to invest onadditional capacity (green, storage, fossil)

« City »

• Strategy: maintain low energy averageprice, avoid peak prices, preserve comfort(limit “shaving”)

• Tactical play: choose local operator or“classical” supplier, invest into energysavings (megawatts)

(Regional

fossil/nuclear)

Supplier

Regulator

City

(Local

fossil/green)

Operator

CityCityCity

CO2 Tax

CO2 TaxGreen

Incentives/

constraints

energy Wholesale

price

« classical »

distribution

of energy

Energy @ dynamic price

Variable demand

Part

I :

Moti

vati

ons


Regulator /

Environment

Environment parameters

• Demand growth

• Oil Price Trend

• Nuclear Growth / Reduction trend

• CO2 tax

Technology Parameters

Systemic Parameters

• Cost of green tech (yearly

trend

• Cost of storage (yearly trend)

• Demand variability

• NegaWatt generation (alpha)

• Peak shaving (beta)

• Market share sensitivity

(gamma)

Operator

City Supplier

Operator’s

customers

Supplier’s

Customers

Open Market

The supplier buys electricity

on the open Market when

demand exceeds capacity, at

a very high price

Yearly

Investments

• Grow / reduce

nuclear assets

• Add fossil capacity

Yearly

• Adjust market

shares• Invest into

« negawatt »

energy saving

equipments

Yearly

Investments • Add « green »

capacity

• Add fossil capacity

• Add storage capacity

ReserveBuffer

Storage is divided into

Two separate units with

Different logics

demand

supply

MM price

o.inBuffer

o.fossilePower

supplydemand

o.sellReserve

o.inReserve

o.greenPower

o.buy

o.sell

o.outBuffer

o.outReserve

Wholesale price

MM price

Simulation

over 15

Years

Systemic Simulation of Smart Grid (Model)


Part II




Part 4: Mass Market Telephony Simulation Example


Game Theoretical Evolutionary Simulation (GTES)

GTES is a tool for looking at a complex

model with too many unknowns

Problem(fuzzy,

complex,

…)

Abstrac-

tions

Model:Set of

equations with

too many

unknown

parameters !

Split

parameters

« Players »

Environment

(DIP)

Strategy (DDP)

Tactical (DV)

“tactical” may

be derived from

“strategy”

(local

optimization)

Parameters which

describe the

player’s goals

eParameters

sParameters

Parameters which are

meaningful (e.g., oil

price future)

Scenario-defined

Obscure &

unknownrandomize

Game Theoretical

Approach

Player’s degrees of

freedom

Wolter

Fabrycky:

DDP/DIP/DV

Part

II :

GT

ES (

Gam

e T

heore

tical

Evolu

tionary

Sim

ula

tion)


Game-Theoretical Evolutionary SimulationPart

II :

GT

ES (

Gam

e T

heore

tical

Evolu

tionary

Sim

ula

tion)

Two ways to look at GTES

Solving a complex undefined optimization problem

Game theory in a complex environment

An approach inspired by Robert Axelrod pioneering work on Agent-based models of cooperation & competition

E.g.; experimental/evolutionary validation of TIT-for-TAT strategy in a repeated prisoner dilemma game

SamplingMonte-Carlo

Search for NashEquilibriums

Machine LearningLocal optimization

GTES parametric analysis

GlobalParameterizedOptimizationProblem

Parameters

Strategy

External

TacticalNon-cooperativeRepeatedGame

Strategic analysis of the player’s goals

Classification

Taking the uncertainty of the model into accountActors

Context


Evolutionary Algorithms & Machine Learning

From each actor’s viewpoint, everything being equal, a GTES model defines a parametric optimization problem:

The objective function fpi for each actor represents the « strategy»

A set of state variables and associated target values (e.g., EBITDA, share, …)

Linear combination + concave valuation of difference

The set x of free variables represent the « tactic »

Finding the best solution (called BR: Best Response) requires to solve an optimization problem

Machine learning means that finding the best tactic is automated

An approximate model requires a heuristic solution

Hill-climbing / meta-heuristics (SA or genetic algorithms)

Local moves according to a neighborhood structure + dichotomy search

Choosing the proper neighborhood structure is a key modeling choice

Eeexf pXx

,,maxMulti-actor maximization

« game/problem » (fp Rn)

Bounded rationality

Part

II :

GT

ES (

Gam

e T

heore

tical

Evolu

tionary

Sim

ula

tion)


The Search for Nash Equilibriums

RAIRO - Operations

Research

Vol. 43 No. 4 (October-

December 2009)

Nash Equilibrium (NE)

When each actor is maximally satisfied, w.r.t. each other actor’s tactic

The simplest way to find a NE is to iterate the computation of the « Best Response » function

An iterative loop that may be nested with the local optimization loop

A heuristic version may be derived according to a neighborhood structure V

There does not necessarily exist a « pure » Nash Equilibrium

The loop may not converge ( “destructive war” or “chaos”)

The convergence rate increases with a « maxmin » approach

The valuation function is extended to take one level of feedback

Hence producing the concept of « Forward looking Nash Equilibrium »

),(),(,, iiiiii ttftxfTxi

),(),(,, **

iiiii ttftxfTxi

)),(,((min),(*

jiViiij

iii tjBRtfttf

Part

II :

GT

ES (

Gam

e T

heore

tical

Evolu

tionary

Sim

ula

tion)


Sampling

Monte-Carlo

The uncertainty regarding the environment parameters

e is handled through randomization

Each parameter from E is drawn between an min/max

Example: a [1.0, 3.0]

Scenarios

Are used to implement « what-if » analysis (though e)

Boundaries for Monte-Carlo sampling

Experiences

Sample Size x Scenario x Strategies

For each sample, we search for a NE through

a fixed number of iterations

Result is a triplet

Classification (% of stable, war, chaos)

Typical values of key “business” status variable

(mean + confidence intervals)

Stability metric (rate of convergence,

standard deviation ratios) 0

1000

2000

3000

4000

5000

6000

7000

1

37

73

109

145

181

217

253

289

325

361

397

433

469

505

541

577

0

1000

2000

3000

4000

5000

6000

7000

8000

9000

1

18

35

52

69

86

103

120

137

154

171

188

205

222

239

256

273

290

0

1000

2000

3000

4000

5000

6000

7000

135

69

103

137

171

205

239

273

307

341

375

409

443

477

511

545

579

Part

II :

GT

ES (

Gam

e T

heore

tical

Evolu

tionary

Sim

ula

tion)

stable

chaos

?


Lessons from Practice

Defining the satisfaction (w.r.t strategy = set of goals) is critical

Additive versus multiplicative formulas

Use multiple strategy objects to avoid the fine tuning of sensitivity

One way to resolve the relative weight issue

Local optimization = Neighborhood + meta-heuristics

Simple local-climbing seems enough …

… but the need for “multiple simultaneous changes” (e.g., 3-opt) is an

indication of the game’s interest

Experiences with meta-heuristics (Tabu, genetic, random walks) are

interesting but do not change the nature of the result

Sensitivity to initial values for “tactics” is a quality indicator of

local search strategy

Meta-principle : increase the “opt power” until stability is reached

Measuring “Nash convergence” is tricky (unless infinite time)

Easy : define a N-uple distance over tactics

Harder : evaluate if a small distance is acceptable

Linear regression on major Business KPIsPart

II :

GT

ES (

Gam

e T

heore

tical

Evolu

tionary

Sim

ula

tion)


Part III





Part 5: Conclusion


Energy demand Market Share

Demand generation

Operator Production

Dynamic Pricing« NegaWatt » compensation (yearly)

Peak « shaving » (hourly)

MW

Time (hourly/daily)

pattern

Random

noise

By

City

%

savings

Electricity sell price

a1

a2

b1

b2

%

cutoff

price

Generates investments

As price rise,

Cities invest in energy

saving

Self-motivated or

operator-controlled

Peak price → partial cutoff

g1

g2

Opera

tor

mark

et

share

Price ratio

(supplier/operator)

Max penetration rate

Price sensitivity

Supplier (wholesale/ customer)

Operator

Price

($)

• local → production price + margin1

• supplier → wholesale price + margin2

Production cost × D

Wholesale base price

Production (GW)

D

Nuclear capacity

+ customer

costs

Use local « green » power

• Green power is intermittent

• The operator controls & monitor all green

production from the city

City Energy Demand (MW)

Use local storage (buffer / reserve)

Use local « fossil » power

Wholesale purchase (Supplier)

Extra

demand

Extra

capacity buffer

resell

Extra

demand

+

• use buffer if full

• Use reserve if (buy) price is high

• Fill reserve is (buy) price is low

• Sell from reserve if (sell) price is

high

• Fossil production is variable

• Fossil production generates CO2 taxes

• Unmet demand is bought wholesale

buffer

reserve

-

+/-

Reserve « threshold »

prices → tactical

parameters (policy)

S3G : A Collection of Simple Models

Difference between constrained /unconstrained demand

Part

III :

Sm

art

Gri

ds

Syst

em

ic S

imula

tion


S3G : Players’ objectives (optimization functions)

Each players tries to optimize three “KPI” (performance indicators)

Using a linear combination

Measuring the difference between current and target value (defines a strategy)

Regulator

To maintain total output (economy = electricity consumed + negaWatt)

To reduce CO2

To keep a balanced budget (subsidies < taxes)

City

To keep average electricity bill as low as possible

To keep the current level of demand-response shaving

To reduce the « feared worst price » = peak price + 5 x anual growth rate.

Supplier (global)

To keep income at current level

To protect market share (80% when simulation starts)

To keep the number of hours when foreign supply is needed to a minimum

Operator (local)

To grow market share (from 20%)

To grow turn-over

To increase income (sales – expenses)

Part

III :

Sm

art

Gri

ds

Syst

em

ic S

imula

tion


(1) Implement S3G Model

The model has successively been implemented (1000 lines of CLAIRE code)

“Rules of Play”

S3G Work Plan

(2) What-if analysis and validation

S3G has been checked though a number of what-if scenarios (both as a

debugging method and a first output)

(3) Machine Learning : « Tactic optimization »

A crude version of “hill climbing / local search” optimization is operational.

More complex methods are required because of pricing structure

(4) Search for Nash equilibrium

This is how we address the question of competition vs. cooperation.

(5) Randomize unknown environment parameters

Full-blown GTES simulation includes a Monte-Carlo sampling of unknown systemic

parameters to asses the robustness of phase (4) : classification. Randomization is

extended to demand generation to study the impact of variability

(6) Search for robust strategies and robust equilibriums

The search for best tactics is extended to take robustness into account.

The analysis of the competition landscape is revised accordingly

(7) Scenario Analysis

The last phase is to decompose the parametric space into relevant scenarios

to address the questions/issues from slide # 1.

Environment parameters

• Demand growth

• Oil Price Trend

• Nuclear Growth

• CO2 tax

Technology Parameters

Systemic Parameters

• Cost of green tech

• Cost of storage

• Demand variability

• NegaWatt generation

• Peak shaving

• Market share

sensitivity

S3G tool

RegulatorSupplier

Operator

City

Reflects one’s

vision of World

economy &

policies

Reflects one’s

confidence in

technology

progress

Reflects one’s

understanding

of energy

ecosystem

S3G : Simulation & Game ProtocolPart

III :

Sm

art

Gri

ds

Syst

em

ic S

imula

tion


Sensitivity to variability

Testing the hypothesis that variability favors local operator

The results show only non-significant improvements for SG operator(small compared to the overall “local resell business” equilibrium !)

Variability “pushes” the system in the “right direction” (favorable to smart grids), but

is a “small scale” change and most of it seems absorbed in the complex loop

interactions

variability

Demand

response

shaving

NegaWatt

Investment

Local vs

centralized

fossile production

Wholesale

price

Operator

EBITDA

Market Share negaWatt Fossile

investment

DR shaving

E1: regular 85,13€ 789M€ 22% 7,08TWh 2MW 14,04%

E2: more

variation87,50€ 711M€ 22% 7,15TWh 4MW 14,87%

E3: local

variation84,39 703M€ 22,56% 6,34 12MW 14,7%

Very sensitive to

environment

parameters

Reaction to price

increase

Fear :

projected 5-

yr price

Part

III :

Sm

art

Gri

ds

Syst

em

ic S

imula

tion


Carbon Tax, Green Power and Storage Cost

Carbon Tax

Does not help: reinforces the advantage of nuclear energy

Carbon tax to Solar subsidy: positive (PV industry ) but marginal

Green Power : still too expensive … (all economic parameters drawn from Web search – orders of magnitude)

Storage Cost Local optimization finds the optimal buffer/reserve ratio & when to buy/sell

Efficiency = average difference (buy/sell) price -> depend on price structure !

Yields a price threshold at [50% to 100%] of wholesale price

Resilience (e.g., Japan) or co-usage (electric car) is not factored in

Wholesale

price

Operator

income

Market

Share

Solar

Investment

Storage

Investment

NegaWatt DR

shaving

Total

CO2

E1

Reference84,12€ 862M€ 21,8% 0MW 0MW 7,3TWh 14,3% 35,8Mt

S3:

CO2 tax93,7€ 1106M€ 19,8% 0MW 0MW 9,37TWh 14,5% 24.7Mt

S3a CO2 tax

+ solar92,3€ 1006M€ 18,8% 1670MW 6,8MW 8,63TWh 14,8% 25.4Mt

S4 = S3a +cheap

storage87.8€ 992M€ 21% 416MW 50MW 7,7TWh 15,25% 29.3Mt

S6 cheap

Solar (100€/MWh)83,57€ 786M€ 21,7% 2461MW 0MW 6,9TWh 14,2% 35.7Mt

Part

III :

Sm

art

Gri

ds

Syst

em

ic S

imula

tion


Smart Grids : Strategy Matrix

S3G is a stable model (no war/ chaos) but ESS convergence is approximate

The strategies of suppliers and operators may be aligned or conflicting

The effect is regulation is very important

0

20

40

60

80

100

120

140

160

1 3 5 7 9 11 13 15 17 19

EBITDA (M€)

price (€/MWh)

Nash distance(%)

0

20

40

60

80

100

120

140

160

1 3 5 7 9 11 13 15 17 19

EBITDA (M€)

price (€/MWh)

Nash distance(%)

• wholesale boundaries

• fixed/variable price structure

Operator:

Soft Strategy

Operator:

Hard Strategy

Supplier :

Soft strategy

Supplier: 11635 M€ @ 79.7€

Operator: 1282 M€ : 19.9% MS

Supplier: 11775 M€ @ 80.7€

Operator: 667 M€ : 21.6% MS

Supplier:

Hard strategy

Supplier: 7119 M€ @ 70.3€

Operator: 1253 M€ : 20.58% MS

Supplier: 7225 M€ @ 68.7€

Operator: 737 M€ : 19.9% MS

Focus on

Marketshare

Surprise ?

Part

III :

Sm

art

Gri

ds

Syst

em

ic S

imula

tion


Oil Price Sensitivity and « De-nuclearization »

Oil Price increase does not favor smart grids operators

De-nuclearization is a more favorable scenario …

When storage cost is lowered (1% market share gain at 100$/W)

Studying scale-sensitivity would require more time/computers/faster

machines

One game with approximate results (10 samples) : 1 day of CPU

Based on other GTES application, typical sample size should be 1000

Wholesale

price

Operator

income

Market

Share

Solar

Investment

Storage

Investment

NegaWatt DR

Shaving

Total

CO2

E1

Reference84,12€ 862M€ 21,8% 0MW 0MW 7,3TWh 14,3% 35,8Mt

E4: oil price

increase91,2€ 647M€ 21,4% 0MW 5MW 8,3TWh 15,5% 28Mt

S2: Government

« de-nuclearizes »88€ 785M€ 21,7% 0M 4MW 8TWh 14,6% 45Mt

H1: 3 cities (vs 10) 82.05€ 866M€ 21.7% 0MW 0MW 6.9TWh 12.67% 36,1Mt

H2: 20 cities 83.6€ 568M€ 23.6% 0MW 0MW 7TWh 13,56% 37Mt

Part

III :

Sm

art

Gri

ds

Syst

em

ic S

imula

tion


S3G Temporary Conclusions

Systemic Simulation of Smart Grids

A very simple model …

… yet which captures a number of interaction loops between players

Demonstrates an interesting level of complexity …

… shown by the “relative difficulty” to get stable Nash Equilibrium

Serious Gaming as a learning tool (not a forecasting tool !)

Takes the various stakeholders viewpoint into account

Build systemic knowledge (understand the environment as a system with feedback loops and delays)

GTES is an interesting approach for serious gaming

A Few Lessons Learned

Importance of regulation

Competition regulation (dynamic & wholesale pricing)

Technology / CO2 incentives

Smart behavior starts with storage

When it becomes affordable at local scale (vs. STEP)

Part

III :

Sm

art

Gri

ds

Syst

em

ic S

imula

tion


Part IV




Part 4: Mass Market Telephony Simulation Examples

Conclusions


Example (1): Distribution Networks

Simple model

Two-steps phases which

are distinct from an

organization viewpoint

Coupling with other operators

through distributors

Serious Games at the excom

level … successful impact

Parc SRO

Année 1Parc BT

Année 1

SOCC

Ne fait rien

Parc SRO

Année 2Parc BT

Année 2

Churne (interne ou externe)

renouvelle

R1 :

RCBT

R2:

WEBTR3:

GSATR4:

DCT

R5:

agences

R6:

WEB

@SRO

100 clients(forfaits) « statistiques »

Année 2

Année 1

R1 R2 R3 R6

Calcul

Renouvellement

Par réseau

Agrégation

Résultat :

- Bilan par

réseau

- Parc total

par opérateurCalcul

Ventes

Par réseau

Base

100.0

« ajustée »Répartition

Courbe « en S »

d’appétence

Courbe « en S » de

renouvellement

Calcul

PdM

BT/ SRO

Calcul

PdM

BT/ SRO

Part

IV

: T

ele

phony S

imula

tion E

xam

ple

s


Example (2) : Commercial Costs Optimization

Problem: resource allocation between different channels …

… while taking competition between distribution channels into

account (hard to evaluate)

Goal (met) : start discussion between channels

method: what-if scenarios

First round : calibration

Second round : global simulation

Adjustment

Optimization

delta

OPEX

2006 data

sales, fixed/

variable costs

delta

CAPEX

result

Euros

Competition

Matrix

Sensibility

Price -> Sales

Par

canal C1 @

p1

Dist :

60%

C2 @ p2

Dist :

100%

C2 @ p2

Dist :

100%

C3 @ p3

Dist :

100%

C[1,2] = 50%

C[2,1] =

50%

Flux si p1 < p2

Part

IV

: T

ele

phony S

imula

tion E

xam

ple

s


Example (3) : CGS – Cellular Game Simulation

Simulation of competition between n mass-market telephony

operators

Mobile telephony, internet access provider, most operator business

Follow-up of UMTS 2000 simulation

While taking distribution channels into account

Make use of only published data (Annual reports)

… which required successive simplification iterations

Time unit = year (3/5 year => 3/5 iterations)

Model that has been played with 3 and 4 players …

First to evaluate 3 years plans (3YP) robustness

Then, to simulate the arrival of Free on the French market

(2009-2010)

Independently, we run a “war game” organized by

McKinsey, with a more sophisticated model but no

automated reaction

Serious game … are

serious : cannot

comment !

Part

IV

: T

ele

phony S

imula

tion E

xam

ple

s


CGS – Operator Model from CEO’s insight

Input

ACQ: average acquisition cost per customer

FID: loyalty cost

“average service basket” price

(PPM: typical average package price)

Internal Variables

Customer Base

# acquisitions, # renewals

Consumption (faction of expected average)

ARPU (summation of PPM x usage)

Simple Financial Model

Operations expenses (including annual trend)

Inbound / Outbound Turnover, interco (TA)

Ebitda = CA – DO – FID – ACQ - Interco

PU

ACQ

renouvellement

1

acquisition

Opérateur

2

churn

3

PU

FIDprix

MVNO inclus

Aggregation:

•MVNO

•Voice / data

• MM / Business

• Pre-/post-paid

Part

IV

: T

ele

phony S

imula

tion E

xam

ple

s


Coupling

Acquisition/ price relationship →

Change / price relationship [churn / renewal] → similar model with

different parameters

1

Orange

2 3 1

SFR

2 3

1

Bouygues

2 3

Nouveaux

S-curve to model sensitivity

+ competition model Cf. INRIA

2010 Talk

1

Opérateur

2 3

Part

IV

: T

ele

phony S

imula

tion E

xam

ple

s


CGS – Simulation Architecture

5-steps computational model

Each iteration produces yearly results (model’s variables) for each

operator

Nouveaux

clients

Calcul Churn,

Renouvellement,

migration

3YP – tactique :

fid, acq, pricing

f,f' : Courbes en S +

compétition (opérateur y

1

Renouvellement

(global –

non ventilé par

canal)

45VolumesRésultats

Canaux

Calcul

Ventes

Par canal

PdM

Par canal

2

3

Base

op 1

Base

op 2

Part

IV

: T

ele

phony S

imula

tion E

xam

ple

s


Open Mass-Market Naive Competition

(Example 4)

Each actor is a company with

an ARPU (price)

an attractiveness (premium)

a customer base

fixed costs + variables costs

migration fluidity

a structural churn

Crude

estimates !

Market share follows a

(price + premium)a distribution

(new)

price

(old)

price

mig

ratio

n

Churn follows a

constant × priceb

distribution

CAVEAT

- closed market

- retail channels ignored

- no segmentation

⇒ 50 lines of code,

easy to reproducem

igra

tion

mig

ratio

n

…

(new)

price

(old)

price

Unknown:

- a

- b

Customer

base 1

Customer

base N

Part

IV

: T

ele

phony S

imula

tion E

xam

ple

s


Mobile Operators (I) : Playing What-If ScenariosPart

IV

: T

ele

phony S

imula

tion E

xam

ple

s

-500

0

500

1000

1500

2000

2500

3000

3500

4000

4500

2011 2012 2013 2014 2015

B

O

S

F

M

3 mobile operators 4 mobile operators

4 mobile operators, loose strategies 4 mobile operators, tight strategies

0%

20%

40%

60%

80%

100%

B O S F M

dev

sat%

result

Stable

WAR

Chaos

-500

0

500

1000

1500

2000

2500

3000

3500

4000

4500

2011 2012 2013 2014 2015

B

O

S

F

M

Stable

WAR

Chaos

0%

20%

40%

60%

80%

100%

B O S F M

dev

sat%

result

-500

0

500

1000

1500

2000

2500

3000

3500

4000

4500

2011 2012 2013 2014 2015

B

O

S

F

M

Stable

WAR

Chaos

0%

20%

40%

60%

80%

100%

B O S F M

dev

sat%

result

0

500

1000

1500

2000

2500

3000

3500

4000

4500

5000

2011 2012 2013 2014 2015

B

O

S

M

0%

20%

40%

60%

80%

100%

B O S M

dev

sat%

result

Stable

WAR

Chaos


Mobile Operators (II) : Strategy Analysis

Sensitivity to alpha (aggressive) Sensitivity to alpha (conservative)

If the fourth operator builds a network ? If the third operator flattens its costs ?

-500

0

500

1000

1500

2000

2500

3000

3500

4000

4500

2011 2012 2013 2014 2015

B

O

S

F

M

Stable

WAR

Chaos

0%

20%

40%

60%

80%

100%

B O S F M

dev

sat%

result

-500

0

500

1000

1500

2000

2500

3000

3500

4000

4500

2011 2012 2013 2014 2015

B

O

S

F

M

Stable

WAR

Chaos

0%

20%

40%

60%

80%

100%

B O S F M

dev

sat%

result

-500

0

500

1000

1500

2000

2500

3000

3500

4000

4500

2011 2012 2013 2014 2015

B

O

S

F

M

Stable

WAR

Chaos

0%

20%

40%

60%

80%

100%

B O S F M

dev

sat%

result

Stable

WAR

Chaos

0%

20%

40%

60%

80%

100%

B O S F M

dev

sat%

result

-500

0

500

1000

1500

2000

2500

3000

3500

4000

4500

2011 2012 2013 2014 2015

B

O

S

F

M

Part

IV

: T

ele

phony S

imula

tion E

xam

ple

s


Part V





Future Directions & Conclusions


Performance Issues

Models that have been tested with GTES are computationally simple, still running once simulation ranges from tens of milliseconds (most of them) to one second.

Tactics’ optimization requires from 100 to 1000 simulation cycles

The search for Nash equilibrium requires many hundreds of optimization cycles (interlacing between the two loops helps by a factor of 10)

Monte-Carlo Sampling requires a few hundreds to a few thousands of Nash equilibriums searches

Consequently,

Computation time quickly becomes a problem(from a day to a year)

Parallel computation is straightforward

Part

V:

Futu

re D

irecti

ons

& C

onclu

sion


Quality of Model Issues

Modeling « Torture Bench »

Machine learning => zooms on logic faults

No mercy for linear approximation that are “locally right”

« Model tuning » takes time !

A good practice is to limit oneself to status variable for which an value history is available

« stability » requirement

Limit / boundary behaviors

Ex: S-curve versus linear formulas

Concavity/convexity –

Neighborhood structure and exploration strategy

Tradeoff between efficiency and relevance (should mimic actors)

Scale –sensitive : need to work on the real size problem (even if abstracted)

Monte-Carlo sampling must be introduced early on to avoid over-engineering

Part

V:

Futu

re D

irecti

ons

& C

onclu

sion


Relations with System Dynamics

System Dynamics

Models based on interactionnetworks between statevariables (e.g., CGS example)

Proximity

Very close to the workof J. Forrester or J. Sterman

These networks are a good first step towardsGTES modesl

Differences

What’s inside the model (detail = interaction formulas) is critical andhas a deep impact on results …

… especially when the system is coupled with a stochastic input flow

Polarities (+/- ) between state variables are not enough, not even the values of local derivatives (elasticity => linear model).

Prix Prix TA

Usage

PU FID

Acquisition

Interco EBITDA

Dépenses

Churn

Renouvellement

PUACQ

123

PUFID

Base

TCOOffre

ARPU

CA Entrant

CA Sortant

PU ACQ

FID

ACQ

prix

Part

V:

Futu

re D

irecti

ons

& C

onclu

sion


Serious Games : Key Take-aways

Results

Game

Analysis

The model is wrong …

or too complex

(e.g. Social Networks)

The model is wrong …

but may be fixed

(e.g. CGS)

The model is right …

our thinking was wrong

Systemic education

The model is right …

and shows a feedback

loop that we missed

The more embarrassing, the more useful

eg: Market Share , Technology introduction …

Many successful instances over the years

Sales Channel, Customer Lifecycle, LTE bid

The expert

disagrees

and the fun

starts

Simple

Model + Key

Business

Variables

What-if Scenarios,

Players’ strategies

Contin

uous

impro

vem

ent

Models’ torture bench

Part

V:

Futu

re D

irecti

ons

& C

onclu

sion


Failed SNS Experiment

I applied GTES to study the “Google+ versus

Facebook” fight two years ago.

Value of SNS experience =

quality of social content X quality of Edge ranking

The model is interesting, it shows the recursive value equation

(strong reinforcement)

Quality of content = f(size of network x time spent on network)

Time spent on SNS = f(Quality of content / effort)

The model was easy to implement (using Duncan Watts

principles for social network growth)

Showcases the difference between adoption and usage !

However, the results are immensely sensitive to frequencies (of visit)

and delays

Bottom-line : GTES is not universal

Start with a few scenarios, then heavy sampling with Monte-Carlo …

If unstable, you need more real-life measures to narrow the

incertitude about systemic parameters

Part

V:

Futu

re D

irecti

ons

& C

onclu

sion


RTMS : Repeated Tender Market Share

Problem: find a model to reproduce the behavior of buyers /

sellers in a closed repeated tender market

For instance, IT division buying software development man days

See if there is a systemic justification for observed practices

First Model

Bid price is a combination (either fixed or randomized) of

Balanced price (economic optimization)

Dynamic price (based on previous bidding history)

Two sets of coefficient according the previous

Selection is based on price (cheapest wins) with a bias towards

diversity (cf. the two-sourcing rules)

Very similar to repeated Prisoner's dilemma game (Axelrod) but more

complex (than TIT-for-TAT)

Preliminary results

Interesting : Nash equilibriums are found … not always (and required

a lot of tuning).

“Forward Nash Equilibrium” – interesting but expensive

Meaningful simulations : collusions, bluffing, …

Part

V:

Futu

re D

irecti

ons

& C

onclu

sion


Future : Connected Health Trust Game

The problem : data privacy issues with connected devices

Would you share your health data with your insurance company to

get a better price ?

Would you react socially - as a group – to selective price increases ?

Insurance issues :

Anti-selection (if another insurer gets the “lower risk” group)

Asymmetry of information

The model:

Segments of population with different health risk behaviors, which

are revealed (or not) by connected devices

Group of insurance companies with different policies (fixed or

variable prices according to behavior)

Macro parameters

Precision of determination - link between behavior & risk

Stability of determination – evolution in time

Importance of social behavior

No results yet … stay tuned (open for collaboration)

Part

V:

Futu

re D

irecti

ons

& C

onclu

sion


Conclusion

“Serious Games” approach will become mainstream in the future

Forecasting does not work any longer

Need to develop reflexes and skills

Build systemic knowledge (understand the environment as a system with feedback loops and delays)

GTES is an interesting platform for serious gaming

Combination of “classical techniques”

Evolutionary game theory … will become popular with faster computers

A workbench for model tuning –CAVEAT – not all models adapt to GTES

Not a panacea, but proven utility over the last 10 years

Still, a lot of work is required

Parallelization (good MapReduce candidate)

Making “Forward Nash” (look-ahead) practical

Leveraging evolutionary meta-heuristics (e.g. genetic algorithms)

« The difficulty lies, not in the

new ideas, but in escaping

from the old ones.”

J. M. Keynes

Part

V:

Futu

re D

irecti

ons

& C

onclu

sion

Science

GTES UTC 2014