The Logic of Models and Interventions, MIke Moldoveanu
Breaking the Code of Change II, Rotman School of Management, August 2-3, 2000
These participant's notes were created in real-time during the meeting, based on the speaker's presentation(s) and comments from the audience. These should not be viewed as official transcripts of the meeting, but only as an interpretation by a single individual. Lapses, grammatical errors, and typing mistakes may not have been corrected. Questions about content should be directed to the originator. These notes have been contributed by David Ing (daviding@systemicbusiness.org) at the IBM Advanced Business Institute ( http://www.ibm.com/abi).
Mihnea Modoveanu, Rotman School
Specializes in people and the way they think about change.
Wants to talk about causal models behind the causal models.
Causal models create mental maps.
May not agree with everything, but easier to ask for foregiveness than permission.
Purpose of the presentation:
Which models are most powerful?
Relevant to managers who can choose.
What should be the criteria for choice on models?
Academics might think empirical.
To guide managers to use models.
To guide academics to create more powerful models.
Won't talk about the proof of theory, or the predictability.
How successful is a model in gripping minds?
Nietzsche, It's not the truth of the idea that matters....
Causal models, to take us from unbearable present --> desired future.
Dorner (1999), The Logic of Failure
Given a simulated African country, where you are free to do everything.
Tupis, animals, farm animals, ...
Top chart: Most simulations result in disasters
Bottom chart:
How have people acted?
What cognitive processes take place?
They make frequent choice not only about what to do, but also the causal models underneath, because they search for new data -- in particular, disconfirming models.
Porter (1996) What is Strategy?
Productivity frontier, value vs. cost of producing value
Choice ratification versus choice implementation
A causal model is series of propositions (which should be consistent).
Induction is a psychological mechanism: the fact that it's happened n-1 out of n times, doesn't mean the causal model is correct.
How to choose a model: 3 criteria:
Controllability: Axioms: want control over independent variables.
Agency theory says have lots of choice, i.e. different types of contracts, can control fate.
Computability: if had Turing machine, would it converge to an answer in a reasonable amount of time?
Observability: Does the model make observable results, i.e. intersubjective agreement.
Then models that are better on all three criteria would motivate more use of the model.
Empirical base
Informed by action science.
Epistemology: Reference theory of truth, building more successful theories.
Cognitive psychology: The way people make decisions.
Sociology of science: What types of models are likely to take hold?
Studies of Decision Agents' Relationship to Own Beliefs: Mike's research.
Studies of venture capitalists' processes of model selection and validation.
Question: Models as consistent with what's already in their head?
Theories are hard to refute.
Theories taking hold of people, as opposed to people taking hold of theories.
Two stage process:
First, a lottery over choices of beliefs.
Second, given that you've accepted these beliefs, then a choice of actions.
Believe change happens in the first stage -- where there's a lens.
Actions are constrained by the first stage.
Computability failures:
Model of testing for computability:
You have a theory, and a Turing machine version of the theory.
Will it converge in a reasonable amount of time?
When I say "I am telling a lie", won't get convergence.
Alternative: a level of practical uncomputability.
I know, you know, I know ...
Primates are trusting creatures, and humans are trusting.
People try to avoid this complexity class.
Game theory answers are therefore not appealing.
Computabiity criteria:
Firstly, next a scratchpad memory restriction of 4 to 7 axioms (e.g. five forces, 7Ss).
Secondly, how hard is the problem to be solved (P-hard).
Three complexity classes:
Polynomial
Exponential (non-polynomial hard).
Uncomputable.
Practical failure of computability:
Case of coupled market shares: e.g. Apple, Compaq and Dell.
Depends on market share evolution R: how much is the network externality? How many people have bought in the past?
If R is above 3, then chaotic market.
Which lead to P problems, and which lead to NP problems?
Managers will choose to use models that solve P-hard problems.
Game theory: Nash is NP hard.
Situational analysis is NP hard.
Consulting industry: may want to farm out NP hard problems (whether or not the individuals can really solve them as NP-hard).
Minimum vertex cover NP-hard: is equivalent to diagnostic analysis, data with hypothesis.
Software design: Are the goals compatible with the constraints?
Complexity.
Nash equilibria are NP hard: No wonder people don't play to Nash equilibrium.
Rules based inference: Consistency checks on finite sets of axioms is NP-hard.
Inference to the best explanation, situation analysis is NP-hard.
Uncomputable problems:
Interactive reasoning means an infinite dimension.
Moral deliberation
Epistemic deliberation: Can I create a set of axioms .... (Hilbert-Godel).
Reflexive equilibrium:
Is there a reflexive equilibrium in the market, e.g. where there is information about information about information ...
Impossible.
Complexity classes:
P-class: sorting, allocation
NP-class: non-linear, constrained.
Careful to don't give people non-computable problems.
Questions
Heuristics cut problems down from NP problems to P problems.
Stochastic problems?
Observability and controllability
Observability: Intersubjectively measurable?
Controllability: Are there some levers that the managers can pull?
Theories on (1) computability (2) controllability and (3) observability:
Traditional Industrial Organization: easy.
Interactive I O Theory: Game theory: low computability
Evolutionary Economics: low on computability and controllability.
Stochastic, mutations.
Go to coin-flipping strategy.
Question: Can NP-hard problems be reduced to P-hard?
Complexity theory.
Robots moving something large, versus ants: ants use simple rules, then coordinate; robot can be programmed to do the same.
Problem is making the models explicit.
(Continue)
Agency theory: high on computability and controllability, but low on observablility
Need to avoid vicious circles:
Starting with a low observability model, work on it to make it less computability, and then tinker to make it less controllable.
Tradeoffs between computability, controllability and observability.
Conclusions:
Cognitive choices matter: Not the only choices, but are part of the choices.
Model as action maps: they turn predicaments into problem statement and problem statements into puzzle statements.
Theories boil down to model
Choose models on higher computability, controllability and observability frontiers.
Question: Can we replace a NP-hard problem with low dimensionality over P-hard high dimensionality problems?
Hypothesis: looking for alternatives.
Response: Reducing to five forces is reducing dimensionality, which made it easier to accept.
Mike Porter: IO theory didn't capture everything that managers knew. Five forces created buckets, both increasing complexity and reducing complexity.
Maybe, but think of this as a theory: not just complexity, but also controllability and observability.
People shy away from unjustifiable choices:
First you choose a model, then have a justification for it.
Discussant: Glen Whyte
Some content on this website may be subject to prior copyrights.
Please contact the author(s) prior to reproduction or further distribution of the materials.