EXYSTENCE NoE
Seminar on 12 Nov 2004, Helsinki

Prof. John Casti, Institute for Monetary Economics, Vienna, Austria and Complexica Inc. Santa Fe, NM, USA.

The World of Business in a Box

Today I think that we have heard a lot of about individual companies. Eve talked about Rolls Royce Marine, Hannu talked about UPM and there was some fairly general discussion involving sociology, philosophy, psychology and some things in between about motivation and understanding about how a business enterprise actually works. The topic that I'm presenting today is not of that kind, though it is perhaps something which focuses on trying to understand business as a science. One of the crucial aspects of any kind of actual science is the ability to do controlled containable experiments and hence to test hypotheses.
In the natural sciences, physics, chemistry and biology to some degree, it has been almost taken for granted from the time of say, Galileo, that you study billiard balls or planets or electrons or elementary particles. You construct laboratories of some kind and use them to test hypotheses by doing repeatable experiments. However it's one of the many sad facts of life, that in the areas of everyday life that we most care about, this kind of experimentation is exactly what you can't do. If you have a theory of finance you cannot go down to Wall Street and ask them to change the rules for the day to test your theory and even if you could, you could never repeat the same experiment. Similarly with road traffic problems, communication networks, health care delivery systems, paper companies and engineering manufacturing firms. The ones that I think are the most important examples of complex adaptive systems (or co-creative) are beyond this kind of experimentation. So it's very difficult in the social behavioral realm to follow the dictates of the scientific method to create something that looks like a decent scientific theory of these processes if you don't have laboratories to test hypotheses. Fortunately, to some degree, technology is starting to come to our rescue by providing us with the kind of cheap powerful computational capabilities to create electronic copies of the systems inside our computers and use the computer as a laboratory for doing experimentation. Roughly speaking, in my view, this is what agent-based models are about. They are about doing the experiment. I want to say to Mika that my hands are not quite as clean as he might believe; they have been and will be again dirty, but I wash them every now and then because when I write books, my keyboard doesn't like dirty hands. But they'll get dirty again and today you'll see an example of that. I'm going to call this talk the 'World of Business in a Box' and it's not a very big box.
The idea is to use the computer as a laboratory, which is more revolutionary than you might imagine philosophically because some of you may have seen a book by Stephan Wolffram called A New Kind of Science. Interestingly enough, this book was almost universally hated by everyone in the complex systems business. I actually wrote the review for Nature magazine and I didn't hate it as much as some of my colleagues in Santa Fe. I thought it had some very redeeming qualities and the most important part is the title, because what I am describing today and what Stephan Wolffram tried using 900 pages was definitely a new kind of science, though it actually goes back a couple of thousand years to the time of Aristotle. Those of you who remember your Aristotelian philosophy will know that Aristotle claimed that if you asked the question 'why?' the answer starts with 'because' and that there were four different ways in which to answer it, now called the causal categories.
The first of the ways, Aristotle claimed, was 'material cause'; for example, my house is the way it is because of the matter of which it is composed- wood, stone etc. The second way of answering the question: 'Why is this house this way?' is 'efficient cause' which roughly is about the energy involved. For example: 'the house looks this way because of the energy the builder has put in constructing it'. These two causes are roughly the way that classical science, especially the natural sciences, answers questions and the laboratories of the natural sciences are focused on exploring the material and the energetic structure of systems.
Of the two additional causes the third was 'formal causation' which has to do with essential information. For example: 'your house is this way because of the blueprint or the plan by which the architect designed this particular house and not a different one'. This is where style starts to enter traditional science on an equal footing because when you use the computer as a laboratory for doing experiments what you're really exploring is not the material, or the energetic structuring of the system, but its information structure; in other words how the information relates one piece to another.
This is what Stephan Wolffram had in mind when he entitled his book A New Kind Of Science and whilst there are many things to complain about in the book the basic idea is very sound. There is a new kind of science emerging and it is bringing into the scientific world view, for the first time in history, the third Aristotelian cause, 'formal causation', on an equal footing with matter and energy. This brings to our attention, the obvious question of Aristotle's fourth category of causation which we now call 'Final Causation'. This has to do with purpose, desire, and will. We would say, for example: 'your house is this way and not some other way because you wanted it that way'. It will be a great challenge; maybe the challenge of the twenty first century, to fold into an overall view of the world, matter, energy, information, purpose, will and desire on an equal footing. I don't know how it's going to be done, but I have a feeling that's it's probably tied up with issues like consciousness. For the moment however, let's stay with third cause-formal causation. I want to tell you a couple of stories and give you a dirty hands example of how this, or complexity science, has actually been used to address some questions in the business world.
First I want to try to separate out the notion of complication from the idea of complexity because very often these two things are confused, especially in everyday language. If you take something that is just complicated, like a mechanical clock, you can take it apart and see the gearwheels and springs and so on, but I don't think you can call it complex.
Now I'll give you an example of a combined complicated and complex system. Many of you will remember the currency crisis a few years ago in Mexico. It propagated out from a Mexican peso crisis through a variety of different countries in South America and to this day, Argentina is suffering dramatically because of the consequences of the process. What it demonstrates is that there are some very important long term effects that explode out of the system if you push it in the wrong way in the wrong place at the wrong time. What it also shows, and I'm sure everyone in his room could dramatically add to this list, is the fact that complex adaptive systems are the systems of everyday life. There are no simple systems. If you deal with real systems in the real world you're dealing with complex systems. The only simple systems are in text books and not very good text books. If you want to come to terms with real problems in the real world then you have to face up to the challenge of what we would call complex adaptive systems.
A familiar real world example is an American football game. I was doing some deep experimentation on Super bowl football games a few years ago, trying to decide how I should have placed my bets for the game. Since there is no good mathematical theory for this kind of process I had to resort to something that I did know, which was to play this game a few thousand times inside my computer and do some analysis on what happened. If you want to know how this experiment came out it's the opening section of Chapter One in a book I wrote a few years ago called Would Be Worlds, which is about computer simulation and agent-based models. I'm very happy to mention it here in Finland because I was going to call it 'Counterfeit Worlds' which I didn't like because it suggested something which was not of true value. One day I was in the Institute for Future Studies that happened to be in Stockholm, not Helsinki, and I was sitting in the coffee room talking to the Director and a Finnish professor came up with three students and said: 'why don't you tell these people about the book you're writing?' I described a bit of what I was trying to do and one of the students said: 'Oh you mean a would-be-world?', and I said: 'That's it, a would-be-world'. OK, here are some of the characteristics that all complex adaptive systems share:

(a) A medium size number of agents making up the system. In the football game it's the players, but it could be traders in a financial market or animals in an evolutionary ecosystem or firms in an industry. It means it's not such a small number that you could work out all the interactions with a back of an envelope calculation and it's not such a big number that you can statistically aggregate it into quantities that tell you anything you really want to know about the system. So what is a medium number? Well it's certainly bigger than two. Think about the 'n body gravitational problem' in physics. We know that when 'n' is only two bodies, say two planets moving in each others gravitational field, then there's some formula that, given the position and velocity you can work out where each will be at some future time. However, as soon as you have three or four bodies then, though in principle you could work it out there's no formula. If we think about classical physics at the other end of the scale; say a litre bottle full of gas with about 1023 molecules moving around then, whilst in principle you could work out the future state, in practice you cannot. Of course we would never try to do that anyway because, since Boltzmann, we know that you can statistically aggregate the molecules to higher level quantities called pressure and temperature because the objects are all homogeneous. We also have a relationship which links them called the ideal gas law: PV = nRT. So 'a medium sized number' is more than two and certainly less than 1023. Thus the definition will take into account many systems that we encounter in everyday life in which the number of agents will be a few dozen to a few hundred thousand.

(b) The second property is being 'intelligent and adaptive'. 'Intelligent' here is meant in a very low level sense as 'objects that make up the system interacting according to rules'. At any given moment they 'decide' what action to take. So if in a traffic system you're a driver then you can decide whether to speed up or slow down. In a real system you're continually monitoring your rules to see whether it's producing results that you're satisfied with. If, for example, you're a currency trader at Citibank and you have a rule on whether to buy or sell or hold, then if you're making money you stay with the rule. If the rule stops working then you change to a different rule or invent a new rule. So, in investigating a system you're always exploring the possible space of rules that agents might be following and that's what is meant by 'intelligent'. The system is adaptive because the agents change to a new rule if they don't like the one they've got. And that's exactly the kind of thing that doesn't happen in the natural sciences. The rules that govern planets and electrons are not changing ones. So you could say that physics is a special case of biology though a lot of physicists like to think it's the other way round. I personally think it's much simpler than biology by several orders of magnitude.

(c) The last characterization is 'local information' which just means that there's no object which knows what everyone else is doing. On the football field for example, a player knows what some of the other players are doing but doesn't know what all of them are doing. And a driver on the motorway will know what drivers in the vicinity are doing but not what others farther away are doing. Sometimes that matters and sometimes it doesn't.

Let me now talk about how to use the computer as a laboratory in order to understand problems in the business world. The motivation or reason for building electronic copies of real world systems is to understand the possible consequences of different options. Here is a list of requirements for constructing effective agent-based models; one that is useful in setting up a real business. It is no use creating a world that nobody understands so here are a lot of conditions that have nothing to do with actual techniques but are practical considerations; the rules if you like, that you have to follow if you want customers. Customers want to see things that are easy to use and don't take for ever. Here is the list:

* The simulation must capture the user's consumer, competitive, economic and regulatory environment.
* It must address the user's business interests.
* It should run on the user's computer system.
* It should be easy to use without a manual.
* The development time should be on a reasonable time scale (e.g. 3 months).
* The run time should be reasonable (e.g. a few minutes).
* It must be cost effective.

For the rest of this talk I want to give you a real world example of how the method is used and this has to do with a problem in the world catastrophe insurance industry. A few years ago I attended a meeting in Bermuda which was convened by a lot of the world catastrophe insurers and re-insurers who are the people who insure the insurers. The kind of cover under consideration is for earthquakes, hurricanes, floods, storms and so on, and these people wanted to know how the science of today could help their industry. At the meeting there were four speakers, myself and three climatologists, and the insurers thought that the single most important thing that science could offer them was a better method for predicting hurricanes. I thought this was interesting because if you had a perfect 100% method for predicting hurricanes it would be the worst thing for the insurance industry because you would be out of business. The very essence of insurance is uncertainty, not certainty and that's what you have to consider. If you know nothing however, you have no way of judging the likelihood of a hurricane of certain magnitude striking some place and that's also bad because you don't have any way of knowing what to charge your customers. So logically there is some optimal level of uncertainty for a healthy business, between total knowledge and complete ignorance. Nobody knows that level, but I suggested that with the technology that was then available, we should be able to construct a copy of the situation inside a computer of the interaction of the customers who buy insurance, the primary insurers who issue the policies and the re-insurers who buy some of the risk.
The basic questions that required an answer were:

* How do the frequency, magnitude, and geographical distribution of natural catastrophes affect the profitability of insurers and re-insurers?
* How is the risk spread?
* What is the risk of different pricing strategies on individual companies and the industry as a whole?
* How do the consumers affect the risk?
* What effect does the availability of capital have on risk?
* What are the conditions for the formation of new companies?
* What are the conditions for bankruptcy?
* What is the effect of marketing strategies?
* What is the effect of the structure of insurance/reinsurance contracts?

Every one of the simulations had to have feedback from people in the industry saying what actually happens in their world. The time scale that we used in a particular run was forty time-steps of three months giving a ten year period. This was a long enough period of time for things like inflation, interest and other things to change which suited the demands of the insurers.
Looking at the equity in the five firms considered we saw that one firm went out of business after three years because of a single big hurricane. The company had taken on lots of risk and the hurricane came too soon. As long as nothing happens the company goes on collecting the premiums, but in the case of this company they hadn't collected enough over the period to sustain the losses. So that's an important consideration. The other major consideration is how best do you invest the money you collect if you don't know when something is going to happen? You might have to get your hands on a big chunk of money really quickly so you don't want to put it into investments which charge a high penalty for taking it out. The company that went broke did not diversify sufficiently and sell enough of its risk to the re-insurers and when the hurricane struck they had not accumulated enough capital. This had the unfortunate result that the people whose houses were destroyed didn't get them rebuilt. The other companies stayed about the same as when they started except for one, which didn't have any major casualties in the time and accumulated large amounts of capital.
The important thing about this kind of simulation is that there are many different scenarios that can be played out so that if you're a manager you can ask the question: 'What action should I take today to not only ensure that my company is going to be a survivor, but a beneficiary?' This particular simulation was used mostly as a training tool to introduce new employees to the kind of things that can happen in certain circumstances, but you should remember that there is no one kind of simulation that fits all and you always need expert input.

Questions: What is the difference between agent-based modelling and the kind you are talking about and how much mathematics is there in the simulations?

John: Well, I tend to think that agent-based modelling is a particular style of simulation because you have the individual objects that make up the system and you have to specify rules of interaction which depend on what the other agents are doing at the time. These rules are expressed mathematically. But in general simulations come in lots of sizes and shapes and colours and some are very mathematical like the ones you do in physics. In Stephan Wolffram's book the simulations are a lot more mathematical than the insurance model which is somewhere in between. There was an agent-based model of the stock market done in Santa Fe which had no mathematics in it. There were just kind of trading rules and not much mathematics of any kind so I think it very much depends on the situation. I don't think it matters very much whether you call it a simulation or an agent-based model. It's just a question of whether it answers the questions you're interested in.

Questioner 2; what is the cost of the simulation you describe?

John: Two hundred thousand dollars in 1996. We did it over a period of one year, but if a commercial company did it now, it might cost two hundred and fifty thousand dollars and you would get it in three months.

Questioner 3: I would like to say something about agent-based models. I am doing research using neural networks, fuzzy logic and evolutionary algorithms or co-evolution simulation. On a daily basis I am faced with the problem of approximation of function because if you have to do modelling of human behaviour it's very complex. My lab was involved in one project which was supported by the E.U. called U.N.I.T.E. which was promoting intelligent technologies and was focused on simulating the clients of a commercial bank. We had a very good relationship with the bank and said we needed to categorise the type of client. The coordinator from the bank set up twenty five or twenty six kinds of client profile because they wanted to create a system which would be able to warn the bank which clients would move or leave the bank. Modelling the behaviour of such a client was not easy but we ended up with a 70 or 80% accuracy of characterisation. But communication with the client is difficult because you are not selling a plug and play system so it is crucial that the company is able and willing to cooperate. So I have two questions: (a) how well does the system you are selling adapt to human behaviour? and (b) how do you measure the quality of it?

John: The modelling of human behaviour depends upon how much of the human behaviour you have to model and that in turn depends on the questions the model is designed to address. So if you're creating a road traffic network you have to understand the psychology of different types of drivers and the demographics; where they live and where they shop, where they work, where their children go to school and so on. But it's not important to know things like the colour of their hair or height. I did an exercise two years ago for a supermarket chain and they wanted to understand the dynamics of the shoppers. They had a lot of data on the demographics, whether young or old for example and they had different kinds of shopping lists. So for that exercise the only thing that really mattered was the kind of shopping list; what was it that they intended to buy when they came to that market place. We had to work out how they moved through the store and where the best place was to put things.
When you ask about the model being adaptive, I'm not sure whether you're asking how well the model itself adjusts or how well the agents inside the model are designed to adapt.

Questioner 3: I mean when the model needs new data input.

John: So you mean when new information is available how does it reconfigure? It depends on the model. Reconfiguration for the insurance model was fed in once a year when the simulation was stopped and the users looked to see how it was doing in mirroring what was happening in the real word. The model didn't do it automatically.
The second question you asked was: 'how do you evaluate this kind of model?', and I have what I call 'can you trust it questions'. If you see in the simulation world something happening which is interesting, to what degree can you believe that the same phenomena will appear in the real world? So what you're asking is whether the phenomena are an intrinsic property of the system or an artefact which stems from peculiar starting conditions or perhaps an unwarranted assumption. This question applies to all models and generally speaking I think there are two ways that we try to address that. The first is obvious: if you have data from the real system under seemingly identical circumstances you ask whether the model is behaving in roughly the same way or whether there is at least a strong family resemblance. The second aspect which is more important, is to bring people who have spent their whole lives living with the real system and you say: 'what do you think is going to happen?' You then turn the simulation on and if something takes place that they didn't expect you ask whether they can explain it. In other words does it make sense? It's a bit like evolutionary biology. You set the simulation going and species come into existence. We don't know what they will be but when they appear we ask how possible they are.

Questioner 4: When you talk about 'would-be-worlds' how do I find which set of rules lead to the best possible answer? There are some ways of modelling that can produce something like rules but only applied to a particular case and general rules would probably misfire in a number of cases.

John: Well, that's a good question and in all the exercises that I have participated in you go to the client. For the insurance project we went to the insurers and said: 'under these circumstances what do you do and why do you do it? It's finding a rule of action. The road traffic data is filled with experiments that people have done on drivers, finding out what they do under different circumstances. There's no substitute for that.


Questioner 4: As an anthropologist I ask people for their rules and then I see them do something else.

John: That's OK. If you see them do something different then you have to adjust to what you see.

Questioner 4: Well there are sometimes rules that produce rules and you need to be close to see how that works.

John: Well I can only state what I said at the beginning, and that is that this is not a universal methodology. If you can't produce meaningful rules than you need to do something else.

Eve: I think part of the problem is that such rules don't the status of unchanging laws and we constantly have to ask whether they are principles guiding the action.

John: Yes perhaps it's better to call them policies.

You have to know who your agents are before you can give them rules. If you're engaged in some modelling exercise like a football game and you say these are the questions that I want to address then you know who the agents are and how they will operate under the rules of the game. The context matters and as soon as you specify the context then it becomes easier to say who the agents are and what constraints they have. They're the objects that interact to produce the properties you want to know about. There's no point in considering agents that are outside the context you're interested in.

Questioner 5: I'm doing a thesis on leadership and my main question is: 'How do you run a successful company and how would you characterise that in terms of agent-based modelling?'

Eve: Someone offered earlier to link this to self organisation and motivation. Would you like to go ahead?

Commentator A on leadership: Well I just want to add a few comments to this question of self organisation and motivation. I have had a long experience in business and in trying to reflect on my experiences I came up with one rule which is the key to long term success, not only in business but in life and that is the moral capital. I think that translates as the amount of honesty and the amount of corruption there is in an organisation and in a country. It is something that can be explained by game theory, because in any organisation you want to create value which means you have to play a 'plus sum' or win/win game with all the participants. Game theory tells us quite clearly that in contrast to 'zero sum' games where withholding information and betrayal are the winning strategies, 'plus sum' games depend for success on honesty, integrity and transparency. It becomes the most important success factor in the development of an organisation. We may be able to buy an abundance of capital, knowledge and expertise but moral capital has to be created in situ. Moreover, in selecting a leader, to me integrity is a top priority.

Commentator B on self-organising technology: If you go to the paradigm of complex systems the technology is more or less pre-coded, which means you have to define everything at the start of building the system. You have to consider all the possible things that might come up with the user in order to design a system that doesn't fall down in use, which is pretty impossible. That's really the present paradigm, but in the coming years we will shift to another paradigm which is to use some kind of self-organising technology which means that we are no longer tied to this kind of recoded knowledge. The systems will be truly adaptive in that they will learn as they go along. Take the example of cellular networks which is my area of expertise. Cellular radio networks are full of hundreds of parameters that the radio engineer has to tune up during use to make the system work in the best possible way. It's a very tedious and skilled task and after every major change to the network you have to do it again. I think we may be able to get out of that in the future by the system itself taking care of optimisation, so maybe we shouldn't use the term 'parameters' at all.

Eve: Could I hear people's view on that?

Commentator C: From my point of view it's artificial intelligence that you're talking about. If you've seen the film Artificial Intelligence it's the creation of a system which learns by gathering and using information, but it's the question of whether to use several agents. If you can decompose your task into a number of sub-tasks which can be easily carried out by smaller entities as agents and you allow them to interact then you can progress the technology that way.

Commentator B: I think we are talking about the same thing. I am not talking about artificial humans, but that kind of decomposed task.

Commentator C: But what is the difference between artificial human and this kind of technology?

Commentator B: Well artificial human is one part of it, but I'm also talking about decentralised learning systems.

Eve: Distributive?

Commentator B: Exactly, so we're talking about the same thing.

Commentator C: Well, Sony is doing this IBO stuff (robot learning dog). The next generation of IBO seems to be that they will launch IBO to hundreds of thousands of families in an interacting reinforcement environment. Each IBO will adapt and gather information which is then sent back to the 'mother' computer where the knowledge is fused and sent back again so there's a radically incremental increase in learning about the world.

Eve: And I think there's an important question concerning what the two of you have been talking about, which is the question of how we learn in a distributive way. We have to ask whether it's in terms of intelligence or knowledge or maybe leadership. Maybe that's the future.