Hello,

I've been following the discussion with interest, as well as reading the Imperial College paper with as much attention as possible. Disaster Science is what I do (my current side project is analysis of business response to COVID-19), and most of my modeling is specific to hazards.

I agree that models should use established social theories and understandings, but one of the things that concerns me about that is that some incredibly persistent ideas - like group panic - are in fact fairly thoroughly debunked in literature that is not native to social simulation but merely adjacent. And they end up with some predictive power not because they're somewhat true but because they are also aberrations from rational behavior.

Stay safe out there,

Eileen

On Sun, Mar 22, 2020 at 10:25 AM Frank Dignum <[log in to unmask]> wrote:

Dear Scott,
Let's keep discussing a bit to keep our minds occupied in a useful way, rather than thinking about all the misery going on right now. (e.g. in Zagreb where there was an earthquake and people fled from there homes to be safe, but right away get more chance to be infected, in refugee camps where conditions will be even worse now than before and who knows what will happen when people there get infected)

I am happy that you relate to the developments in physics and economics about abstract models.
It shows that abstract models can be very useful and also very un-useful.
I like the developments in physics as in theoretical physics the researchers mainly develop models based on intuitions. For these models they introduce imaginary particles, concepts and rules that seem (for them!) intuitive to explain different phenomena. None of these things have themselves ever been seen in reality. They only make hypotheses that if this model is correct we should see something happening in some huge experiment as a derivative. But the physicists did not accumulate experiments on dropping apples, pears, chairs, tables, etc. to get to a theory on relativity. Also relativity theory is not very useful to predict whether a chair will break when falling from a certain height. (although in principle it might do this).
When we start generalizing over experiments we might lack the right vocabulary to express the abstract model. We might have to introduce new constructs that are used to link observable parameters in new ways (not directly observable, but at least consistent with observations in particular cases).
I think that is something that economists have never done properly. For them people are a function in the model. And if people in reality would behave like this function all would be well. Unfortunately, people are not functions like that. They interact and are influenced by other's behavior. And they can reflect and change behavior if they think it appropriate (if government wants to limit my autonomy, I will show them that they cannot just do that). So, as long as this fact is ignored in the economic models they will not be very accurate or useful for many purposes.

I am arguing for developing abstract models for social simulation that should be checked in one hand against existing social theories that have been established and used/validated before. And in the other hand should proof their value as being able to generate specific simulation models in contexts that provide useful results.
As an example, for the current simulation we are building for the virus crisis we use a model of needs and values, of norms and practices and of groups. (it is more detailed than this, but this serves as illustration of the point only). Based on these components, that are based on existing theories, we are building a simulation that captures many of the phenomena that we see in the current situation. So, if this simulation proves to be useful it is one of points of reference that shows our abstract concepts are useful. It does not mean that they are the only way of doing things or that they are the best, but just that they proved useful. And just like physicists do we can subsequently put it to test in other cases, debate about the concepts, etc. I would welcome such discussions in the community as it will advance the field and the theory(!) of social simulation.

Just an explicit disclaimer: I just use things from my own experience above as it is easy to relate them for me. I am not claiming to be the only one doing this or the only good one. Many people have contributed to the advancement of the field, but unfortunately a lot of their achievements have not been consolidated in a general accepted framework where it is easily accessible and extendable. It would be great if we could work on that!

Cheers,
Frank.

On 2020-03-22 14:00, Scott Moss wrote:

Dear Frank and everyone,

First (and before reverting to type) I’d like to balance my criticisms with something positive.  I was my very good fortune to discover and to be welcomed by the European social simulation community in the 1990s.  We all had some pretty fundamental professional disagreements but we also had warm, often close, friendships which have lasted over the years.  These friendships have also been professionally fruitful since they supported ongoing and constructively robust argument.  I hope that that will continue to be the case.

Now about theory and evidence.

Nigel Gilbert said, “My model is my theory.”  If we design our models around evidence, then on the Gilbertian dictum we have a genuinely evidence-based theory.  If we have lots of evidence-based models, we can look for commonalities in order to formulate what amount to more general theories.

Evidence shows that models that are built around theories, are too abstract and remote from primary evidence to be useful.  The clearest and most pernicious example is my longstanding bête noir, economics.  As those familiar with the nature and implications of complexity will know, social interaction gives rise to unpredictable events such as economic slumps and, currently relevant, contagion.  Van Parunak pointed out recently on this list that the spread of COVID-19 is likely to be power-law distributed rather than exponential.  That is what we would expect when people behave as social psychologists have repeatedly demonstrated us to behave for nearly a hundred years now.  This means two things:  first we cannot treat the frequency of infection rates as being drawn from a finite-variance distribution so the probabilities cannot be assumed to sum to 1,  Classical statistics and econometrics are therefore not applicable.  Models capturing this realistic and well documented form of social interaction produce outputs that are wholly inconsistent with any form of equilibrium.  Since economic theory is constructed around the concept of equilibrium and econometrics is a development and application of classical statistical theory, both are very poor guides to action of any sort.  Both, explicitly or implicitly, require there to be no direct interaction amongst individuals.  I believe this explains why economic forecasts are systematically (always?) wrong.

I imagine that seeking to impose a unified modelling design, approach, software or whatever top-down, could lead to the same kind of nonsense we have had from economics.  We could start by looking at already published models that are based directly and transparently on evidence to see what is common amongst them.  That is, we can seek to generalise the individual, evidence-based models-as-theories in order to produce more general theories.  Maybe or maybe not one overarching theory.  At an early stage, I imagine there would be a common design spine capturing well documented behavioural norms.

An historical precedent is the development of quantum physics.  Theories were developed to explain observed phenomena.  Some theories implied specific frequency distributions of energy release.  The boson is so called because it exhibits the Bose-Einstein distribution and the fermion because it exhibits the Fermi distribution.  Both of these particles were predicted by theoretical developments of previous theories that were based on observation.

I found over the years that my models were producing power-law distributed aggregate statistics.  I always then found corresponding and independent fine-grained data which also exhibited power-law distributions.  My then colleague, Bruce Edmonds, then found that statistical physicists were finding the same phenomena which further led me to the literature on the conditions in which power-law distributions would be generated.  They were analytically equivalent to the nature of social relations identified by social psychologists from the 1920s onward.

I suspect that validation of social simulation models could well take the form of comparing the parameters of power-law distributions obtained from simulation experiments with the parameters of corresponding real-world power-law distributions.

At my age, I must leave that to you.  I will just continue to goad you in my own cantankerous way.

Thank you for putting up with me (if you have).

Scott

On 22 Mar 2020, at 13:09, Frank Dignum <[log in to unmask]> wrote:

A final remark as reaction to the post of Frederik. I am very happy to see these more thoughtful posts appearing now. It shows that we might not all agree on everything, but there is a fertile ground for some joint development. 
I want to particularly pick on the development of some common ground for social simulations. I believe that this partly should be based on some common understanding on how social sciences theories can be described (formally) and combined(!) for social simulation purposes. Often where theories seem to contradict each other, they actually just use a different perspective because they have different purposes. Capturing these things necessitates careful explorations that are difficult and are often not very much appreciated. I know that there also have been many papers that are just about theory without any connection to reality. However, one can do this theoretical work and explain why you do this with respect to practice. 
In my group we try to do some of this work in order to define some modules that could be used in platforms in an easy way in order to cover some social aspects like norms, practices, culture, etc.
Taking this approach also means that we should take better care of a methodology to use all these concepts. The choice of which concepts one uses for a particular case is crucial. I have done this several times based on my gut feeling and it seemed with good success. But I have no idea if it was the best choice or even how I made that choice.
So, there is lots of work to do in this direction! (which is good because I am not retired yet AND the world could do with more useful social simulations).

Professor Scott Moss
Brookfold
The Wash
Chapel en le Frith
High Peak
SK23 0QW
United Kingdom

tel: +44 (0)1663 750913
mobile: +44 (0)776 968 9991
www: www.scott.moss.name

-- 
Med vänlig hälsning/Best regards,
 
Frank Dignum                    *
Professor Socially Aware AI     *
Department of Computer Science  *
Umeå University                 * 
Sverige                         * Knowledge is only one point, 
e-mail: [log in to unmask]     * the ignorant have multiplied it               
telephone: +46-90-7869101       *                    

---

To unsubscribe from the SIMSOC list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=SIMSOC&A=1

--

Eileen Young

Graduate Research Assistant, Disaster Research Center

PhD. Student, Disaster Science and Management

166 Graham Hall • Newark, DE 19716

---

To unsubscribe from the SIMSOC list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=SIMSOC&A=1