Recent explosive growth in the availability of cheap, powerful and accessible IT capabilities is making it possible, for the first time, to create surrogate versions of real complex systems inside our computers. These 'artificial worlds' will be realistic only if they can simulate the essence of any complex system: the interactions among all its parts, with the system's overall behaviour emerging from these interactions. This cannot be done by traditional analysis-by-decomposition, for example by studying the human immune system or a stock market by breaking the whole into individual parts ¾ molecules or traders ¾ and looking at what these components do in isolation. Maintaining the interactions of the complete system is crucial to the creation of complex adaptive models.

In this report, Professor John Casti describes the opportunities these models offer to create the first real theory of complex systems by enabling controlled, repeatable experiments to be conducted using artificial worlds constructed from informational structures inside computers. He illustrates the theory and practice of artificial words primarily through a work-in-progress report on 'Insurance World', which is being developed for an international consortium of companies in the catastrophe-insurance industry. He describes the theoretical principles and practical rules underpinning Insurance World, as well as highlighting the important practical processes involved in developing such models and the kinds of uses and questions that can be addressed through them. The concepts explained by Professor Casti also form the basis of the SimStore supermarket model, which is the subject of another LSE Complexity Report (Venables and Bilge 1998).

Professor John L. Casti became a resident researcher at the Santa Fe Institute, New Mexico in 1992. His career has also included being one of the first research staff members, in 1974, at the International Institute for Applied Systems Analysis (IIASA) in Vienna; Professor of the Technical University of Vienna's Institute for Econometrics, Operations Research and Systems Theory; and periods of work at the Rand Corporation and University of Arizona. He obtained a Doctorate in Mathematics from the University of Southern California in 1970 and has subsequently studied applied systems modelling in a vast range of activities, from atmospheric radiation to urban housing distribution. Professor Casti's current research interests centre on the way biological metaphors can help to model economic and social phenomena by assisting to establish a coherent theory that incorporates characteristic features of living systems into the Newtonian framework generally used to model non-living phenomena. He is the Executive Editor of the Journal "Complexity". His many publications include key works on complexity (see Bibliography) and a fictional account of a meeting at Cambridge University where J. B. S. Haldane, SchrØ dinger, C. P. Snow, Turing and Wittgenstein discuss thinking machines (Casti 1998).

This report was edited by London-based Editorial Consultant Malcolm Peltu.

Towards a science of social and behavioural phenomena

This report focuses on work at the Santa Fe Institute (SFI) which seeks to understand and model the type of Complex Adaptive System (CAS) found in everyday life. A crucial distinguishing characteristic of such systems is that their component elements are living 'agents' capable of autonomous behaviour, which can be adapted to changing circumstances. This contrasts with complex systems in chemistry, physics and engineering founded on established theories explaining observable phenomena encompassing interactions between non-living elements.

An overwhelming spectrum of living systems falls within the scope of our research, from stock markets to supermarkets (Venables and Bilge 1998), urban traffic networks to prehistoric village formations, national economies to global ecosystems, business organisations to insect colonies. However, although many effective models have been created (Casti 1997), there is still no real science to provide theoretical foundations for building these kinds of systems. One of the main reasons for this slow progress is that researchers into social and behavioural phenomena have not had the ability to conduct the controlled, repeatable experiments that are an integral part of the methods employed in natural sciences to test hypotheses and establish new theories.

Until the advent of widespread and usable computer power, it was generally impractical to perform experiments on everyday social and behavioural systems. For example, Wall Street cannot be asked to change its rules to allow an economist to check a new theory of financial markets. And even if such an unlikely event happened, a genuinely repeatable experiment could not be conducted because too many variables would have to be considered. In other cases, experiments would take too long to be of practical value or may pose too much danger to the real world, say by trying to evaluate a theory about biological diversity by introducing a new species to an environment.

The power and versatility of computer technology has now reached the point where we can create realistic 'silicon surrogates', encapsulating inside a computer the full scope and richness of interactive patterns of the social systems we want to experiment with.

The fingerprints of a complex adaptive system

Research at SFI (Casti 1997) has indicated that the existence of a medium-sized number of intelligent adaptive agents making decisions on the basis of local information can be regarded as the 'fingerprint' indicating that a system being studied can be classified as a CAS. However, these features do not constitute a full definition of complex adaptive systems.

The three distinguishing characteristics of a CAS fingerprint involve:

1. A medium-sized number of individual agents. An agent is the basic element in a CAS. A customer in a supermarket, or a company in an insurance industry, are examples of potential agents in artificial industry worlds. The number of agents must be neither so small that all their interactions could be worked out 'on the back of an envelope', nor so large that statistical aggregation methods could tell you everything you want to know about the system. For example, physics theory enables relationships between the 1023 molecules of gas in a bottle to be statistically aggregated into high-level quantities, like pressure and temperature. These provide valuable information that doesn't rely on keeping track of each of the molecules. In the type of CAS we are concerned with, this isn't the case. The actual number of agents which can be considered 'medium grained' will vary between systems, but a rough guide is something in the range of a few hundred to a few hundred thousand.

2. Intelligent agents with the ability to adapt. Agents need to be 'intelligent' in the sense that they can use in-built rules to decide what actions to take at any given moment. If they find a current rule isn't working well, agents should be 'adaptive' in their ability to discover and change to new or different rules.

3. Local information only. All agents invoke their rules to make decisions on the basis of only partial, or 'local', information. This means there is no agent within the system which knows what every other agent is doing. The 'localness' can relate to physical or informational dimensions. An example of physical locality is the way car drivers can see what motorists in the immediate vicinity are doing, but not what is happening on the other side of town ¾ although distant events could percolate through to affect local traffic dynamics. Informational spaces are typified by the use of a global IT network that can make traders in New York more aware of what is being done by their counterparts in Hong Kong and London than by a colleague at the next desk.

A sporting example of a complex adaptive system

An excellent illustration of a system exhibiting the fingerprints of a CAS is American football. In a simulated 'football world', the agents would be the players on the field. Each player-agent has access to a certain set of rules that can be invoked to decide how a given position should be played. Adaptation is possible because an agent can invent a new rule to deal with each situation; successful adaptation is the main characteristic which separates superstars from the also-rans. No agent can know what all other players are doing at a specific moment, so act on the basis of local information about players in their immediate vicinity.

A coach in the stands might have a broader overview than any player. But the coach is an agent outside the on-field system where actions count. The coach cannot communicate quickly enough, even by radio, to get players to alter their decisions made in real-time while the game is underway, like deciding which way to move next. A traffic observer in a helicopter has a similar relation to drivers in a road traffic system.

Out of personal interest, I have conducted experiments with a simulation of a particular game, the Super Bowl played in January 1995 between the San Francisco 49ers and San Diego Chargers (Casti 1997). This was a detailed micro-simulation based on a codification of the playing records of actual players. One of the aspects of this game which interested me was the almost unbelievable odds on a 49ers win offered by Las Vegas bookmaker. The San Francisco team were quoted as 19-point favourites, which means the bookmakers would pay out only if the 49ers won by more than 19 points. I felt these odds were so great that it would be interesting to see whether they were justified. I used the artificial football world in my computer to play the game over a hundred times, which showed that the 49ers superiority was far less than the bookmakers' 19-point spread. Playing a game a hundred times is impossible in the real world, but easy in a surrogate world in a computer.

A host of other would-be worlds

An artificial American Football World is just one of many CAS simulations which I describe in my book Would-be Worlds (Casti 1997). Others include the TRANSIMS road traffic simulation and an artificial stock market.

TRANSIMS was developed at the Los Alamos National Laboratory to simulate the entire road traffic system of the city of Albuquerque in New Mexico. It is designed to study short-term traffic flows over a period of a few hours, rather than days or months. TRANSIMS includes a vast amount of detail on all the factors affecting traffic, including traffic-light patterns and one-way streets; the location of all roads, small alleys, and residential and commercial buildings; and data on individuals, such as who lives where, what cars people drive, where they live and shop, how many children they have, and where the children go to school. The substantial amount of human and computing resources needed to implement such a large system cost $50 million over five years, with funding coming from the US Department of Transportation. It has enabled many what-if scenarios to be explored in a way that would be inconceivable with the actual traffic system, for example by simulating the effects on traffic movements of changing to a one-way flow on a major thoroughfare in the morning.

The stock-market-inside-a-computer started out with the more academic purpose of studying what is known as the 'efficient market hypothesis', the bedrock of academic theories of finance which assumes complete market efficiency. However, this theory is not believed by anyone working in a real market because there would be no market if it were literally true. Brian Arthur, an economist at SFI, and computer scientist John Holland of the University of Michigan, decided to build a CAS model to help investigate this paradox of a widely held theory which no one in the real world believes. Together with a real-world market trader, financial theorist, and physicist, they built a stock market simulation in which traders are intelligent agents who have rules at their disposal about the regulation of exchanges and the external economic environment.

By manipulating traders' strategies and other factors in their artificial world, Arthur and colleagues have conducted experiments to study the detailed behaviour of actual stock markets. A key overall aim is to discover whether the various adjustable knobs on the model can be set to create circumstances where the efficient market hypothesis is even a defensible approximation to the truth.

Trying to develop a generic CAS model for all industries is unlikely to bear fruit. It is usually best to produce tools honed to doing a specific job well, rather than one that does many things but none particularly well ¾ as happens with machines that try to combine photocopier, fax, laser printer and other capabilities. We are, therefore, mainly interested in creating would-be worlds for specific industries or activities.

My use of the term 'would-be' worlds should not be taken as implying that, as a modeller, I think computer surrogates are somehow preferable to the real world. 'Could-be Worlds' is an equally valid title for my book. My original title was actually 'Counterfeit Worlds'. But at a seminar I was giving at the Stockholm Institute, a Finnish student responded to my description of CAS models by saying enthusiastically: 'Gee, you're talking about would-be worlds'. And that struck me as being a good title for the book.

Applying rule-based models to social systems

Some people are concerned that a rule-based model might be an inappropriate way of expressing the behaviour of systems involving a great deal of human and social interaction. I feel strongly that it would be wrong to exclude many important complex systems from the possibility of being explored in simulated worlds just because people are part of a system's complex interactions.

Its perhaps because I am a mathematician and modeller, rather than a social scientist, that I disagree with the belief that it is impossible to develop any kind of theory about social systems because people are too complex and unfathomable to be encapsulated in equations and rules. The power and versatility of modern computers now makes it possible to use artificial worlds to test out theories about social and behavioural systems which will enable us to get closer to identifying what rules actually drive the relationships between human agents and other system entities.

From a philosophical perspective, it is reasonable to postulate that the behaviour of human agents in a system can be interpreted as being generated by rules, even if the people do not explicitly invoke rules when they act and find it difficult to define such rules if you ask them. It is usually possible to concoct rules to generate particular behaviour patterns, then to experiment with changing and refining the rules to make the generated behaviour more realistic. Excluding the possibility of being able to establish such rules would also make it impractical to even attempt to program computer simulations that could help us to understand complex and critically important social systems.

In some cases, it has indeed been relatively easy to identify a reasonably simple set of rules to generate realistic surrogates of social systems. For example, Arthur and colleagues drew on many books for their stock market model. These contain myriad rules used by traders to decide whether to buy, sell or hold, like the 'moving average' and 'Fibonacci sequence' rules. Other systems can draw on some known, established rules that might not be obvious at first sight, such as those concerning the behaviour of motorists in a road traffic system. There is a substantial road traffic engineering and experimental psychology literature that partly focuses on people's behaviour when driving cars. This research has resulted in much empirical evidence and many formulae that can assist to provide rules for developing systems like TRANSIMS.

For other systems, it is much harder to boil things down to a manageable number of rules. No methods are available at present to determine into which category a given system will fit before you start doing experiments with a simulation to test ideas about how it functions. The exploration of what these rules could be is an important and exciting part of the creation of a model, as is happening in the development of Insurance World (which I discuss later in this report) and J Sainsbury's SimStore supermarket CAS (Venables and Bilge 1998).

Designing experiments with complex adaptive models

As in any other branch of science, the design and construction of complex adaptive models does not start with the experiment itself, but with observations and theory based in those observations. An experiment can then be set up to check up on the coherency of the theory. Results from the experiment are fed back into changes to the theory, which leads to an interactive process of further experimentation and theoretical analysis until, hopefully, there is a convergence towards a solid theoretical foundation. For a CAS model, the definition of agent rules constitute an important part of the theory.

CAS models have been built successfully even when there was not an off-the-peg set of equations and rules to provide the foundation. However, I want to emphasise that there is currently no satisfactory mathematical structures into which the concepts of complex adaptive systems can be accommodated. CAS model builders are in a similar position to gamblers in the Seventeenth century before Fermat and Pascal discovered Probability Theory. It may now be possible to pick a module off a shelf to find out about probabilities in dice, roulette and other gambling games. But before the appropriate theory had been defined, there was no coherent framework for even capturing the kinds of questions we want to ask about CAS behaviour, let alone to provide the answers.

When I made a remark like this at a meeting of the consortium building Insurance World, a manager from a member company asked me: 'But isn't the model's program the theory?'. In a technical sense it is, because a computer program is a mathematical object. However, mathematicians aren't able to gain an insight into a problem by reading the program code. Yet, by developing actual models, running simulations and discussing results of experiments in the artificial laboratories inside our computers, we are gaining insights and building knowledge that is helping to move towards important new theoretical understandings, as well as delivering some tools of great practical value,

The Insurance World consortium

In October 1997, the first meeting was held in London of the Insurance World consortium, which involves eight companies working with SFI to develop a CAS that simulates the catastrophe insurance industry. The companies in the consortium represent a cross section of the large and small firms around the world which offer insurance policies to protect property against natural disasters, like hurricanes, floods and earthquakes.

The consortium came together in response to a talk I gave which indicated the broad potential for studying the dynamics of their operations by creating a computer model of the interplay of factors which make up their industry. I had explained how a CAS model could act as a laboratory for conducting experiments that would be inconceivable in reality, such as seeing the effects on the industry of a Force 5 hurricane hitting Miami beach at a various points in time. The way the industry restructures itself as it evolves according to different patterns of catastrophes could also be analysed, for example by exploring why some companies go bankrupt while others survive, and how introducing new kinds of insurance products can help existing and new companies to prosper.

Traditionally, these insurance firms have devoted a lot of effort to predicting when catastrophes will occur. However, if there were a perfect method of predicting a catastrophe like a hurricane, the industry itself would disappear because people would know when catastrophes would hit. On the other hand, if companies had no knowledge about the comparative likelihood of different events occurring, they would have no rationale for deciding how to place and guard their bets. This suggests there is an optimal level between complete certainty and total ignorance at which the industry is most healthy. I believe a CAS model could help insurance firms to get a feel for this level, as a crucial part of gaining a better understanding of the nature of their business and how to make more money in it. That could be of more value to them than continuing to act as if they were essentially in the business of predicting natural hazards.

The companies which could see the potential advantage of this kind of electronic picture of their industry decided it was worth investing their financial and intellectual resources into a consortium to build Insurance World.

Starting to build a new artificial world

The CAS researchers at SFI who are constructing the model are not experts in the catastrophe insurance industry. So, we stipulated that all consortium members had to be prepared to participate in the entire development process. This was particularly important because we had no explicit set of known rules to draw on to guide the design of a simulation of the catastrophe insurance world. A considerable part of the initial effort would, therefore, involve searching for a set of agreed rules through close collaboration between insurance specialists working in companies in different parts of the world. At the same time, we realised it is important to have an initial model to indicate the kinds of things which could be done as a tangible focal point for consortium meeting.

These considerations led us to plan for an initial one year Phase 1 project. We could scale up this model to industrial strength in a second phase. During Phase 1, five meetings would be held in different parts of the world. This would give member companies the opportunity to critique developments as they unfolded, and to make mid-course corrections to keep the simulator on the right course. The deliverable at the end of Phase 1 in September 1998 would be what we refer to as a 'toy' because it will have very limited basic dimensions, such as the types of disasters which can occur and the number of companies involved. However, all the essential elements on the world, and all interactions between them, would be in place. Although we call it a 'toy', we expect it to demonstrate by the end of the year that it can answer real questions of concern to the firms in the consortium.

This kind of interactive, collaborate bottom-up process of model design, development and experimentation is different to the way some scientific activities might proceed. But it is an essential part of the process of venturing into this kind of unknown complex systems territory.

The basic ingredients of Insurance World

For the remainder of this paper, I would like to discuss how Insurance World exemplifies key principles and uses of CAS models. The capabilities I refer to in the report must be treated as a work-in-progress snapshot of the Phase 1 model, and consortium thinking, at the time of this LSE Complexity Seminar in May 1998.

Insurance World consists of three main conceptual levels:

1. The properties to be insured against the occurrence of certain natural hazardous events.

2. Primary insurers who offer cover to property owners against disastrous events. There are only five active primary insurers in Insurance World Phase 1, compared to a few hundred in the real world.

3. Reinsurers who provide back up to primary insurers. This is analogous to the way in which local bookmakers who get too much action on a particular side of a bet lay off some of their commitment with a larger bookmaker, who becomes their 'reinsurer'. Phase 1 of our model has the capability to have up to five reinsurers, although only two can be active in initial versions. In the real world, there are about twenty to thirty big reinsurers.

There are no 're-reinsurers' in the model or the real world, although a national government might in some circumstances act as a kind of 'reinsurer of last resort'. But government can't, and doesn't want to be, part of the catastrophe reinsurance industry. Some very big companies might also undertake internal re-insurance deals between different divisions. That isn't allowed in the Phase 1 model, but such a capability could be introduced in a future scaled-up version.

The other significant limitations of the model compared to the real world are:

1. only three geographic regions where insurable properties can exist: Japan, California and the Gulf Coast;

2. only two kinds of catastrophe that can occur: earthquakes and hurricanes;

3. restrictions on where events can happen: Japan can experience hurricanes and earthquakes; California has only earthquakes; the Gulf Coast experiences only hurricanes.

To scale up these Phase 1 restrictions to something more realistic, everything except the types of disaster would need to be multiplied by about ten.

We decided not to try to model what is going on within the multiple organisational levels of each firm in Insurance World. Our main purpose is to investigate the macro structures of interaction in a whole industry, not the microstructures of firms within it. This means our energy and thinking has gone into studying interactions between company-agents, where each agent is dealt with as a rule-coded box into which various inputs go and decisions come out. The organisational mechanisms which produce the decisions are not part of our current study. A model giving more emphasis to internal organisational structures would also have been feasible, if that had been what members of the consortium had told us they most wanted to get out of the project within its financial and time constraints.

Rules underpinning Insurance World simulations

In addition to restrictions on the types of catastrophes which can occur in Insurance World and how they can be combined in different regions, the phase 1 model includes three other rules relating to its external environment:

1. The probability per unit time of an earthquake is constant, but is seasonally adjusted for hurricanes, which are concentrated in the autumn.

2. Each catastrophe can occur with low, medium or high intensity, such as having a small earthquake or major hurricane.

3. Investments can be made as either short-term or long-term bonds, both at a constant rate of interest. Insurers and reinsurers generally put part of the premiums they received into investments allowing immediate access in case they have to make major payouts. They use other income to earn a higher rate of interest in longer-term investments that are more difficult to access. The real world has much greater variation in rates of returns and levels of access.

4. The inflation rate for the simulation period can be specified.

We initially kept the interfaces between companies and their external environment to a minimum. However, we are talking to a group in Vienna studying the socio-economic, political and ecological aspects of the catastrophe insurance industry to see if we can introduce other external dimensions based on their research. One of the arts of modelling is to be able to sense where it is best to establish boundaries between the world being simulated and everything else.

Many other rules have been defined to reflect, in a simplified form, factors in the real world that significantly influence interactions between the players in the model: policy holders, primary insurers, and reinsurers. Users of the model can adjust many parametric values to create the operating conditions of the scenario to be simulated. Examples of these in Phase 1 include:

For property covered by insurance policies:

1. average policy amounts for each region and catastrophe;

2. appreciation/depreciation rates for a property;

3. number and rate of growth of insured properties for a region;

4. fraction of properties in a region that will be affected by each type of catastrophe;

5. probability that damage will occur for a catastrophe of a given intensity.

For primary insurers:

1. capital available to a company, consisting of its net worth and outstanding loans;

2. the reserve requirements determining the amount of capital kept by a company in relation to the policies held;

3. maximum amount of loans available to companies to satisfy their reserve requirements, set as a fraction of net worth;

4. the rate, above some threshold, at which loans are paid back;

5. percentage of a company's net worth above the reserve requirement to be paid out periodically as dividends.

For reinsurers:

1. the degree of effort put into increasing market share in a region;

2. the need for reinsurers to cover all excess exposure of primary insurers;

3. no ability for reinsurers to insure themselves (no re-reinsurers).

Insurance World simulation case studies

Insurance World simulations takes place over a ten year period (2000 to 2009). This is divided into 3-month time steps, so there is a total of forty quarters. Many variable parameters might need to be set before a simulation run, for example to establish overall rates for inflation and property appreciation and depreciation, as well as various factors specific to individual companies, like their capital bases. The numbers, timing and locations of disasters can also be specified. Results from simulations are typically presented in graphical forms, such as bar charts showing changing profiles per quarter for the number of insured properties or the market shares in each region.

We have used the Phase 1 model to experiment with a number of scenarios. The following sections summarise the results of a few of these to indicate the types of issues which can be explored. These examples have no more than one catastrophe occurring in a period, but we have run simulations with multiple disasters. In the version used for these case studies, adaptive behaviour was not permitted. This meant managers of agent-companies in the model essentially placed their bets at the outset, then waited to see the outcome after ten simulated years.

Case 0: No catastrophes. Identical companies. Identical initial conditions

In order to test the internal logic of the model, we started with the simplest of scenarios: no disasters occur and companies start out on equal terms. The outcome for this is easy to predict, so we could check that the expected real outcome happened in the artificial world too. The simulation showed a small initial increase in the amount of insurance coverage, caused mainly by the coming onto the market of new properties needing insurance for the first time. As time went by with no hurricanes or earthquakes, property owners began to believe it was unlikely that such a disaster would hit them and so began to cut their insurance coverage. By the end of the ten-tear simulation, the number of insured properties had dropped to a third of the total at the start. As the circumstance were shared equally between all regions, the share of properties insured remained unchanged between regions. This was completely predictable, so validated the basic operations of the model.

Case 1: As for Case 0, but with small variations in initial market shares

We then looked at what would happen if there were small initial variations in the market shares of each primary insurance company, while still keeping the companies identical and having no catastrophes in the ten-year simulation. Here, the company starting with a small edge ahead of its rivals used its differential to dominate the market. This phenomenon is frequently found in chaotic systems, where small changes in initial conditions can create positive feedback that moves things towards a completely different kind of attractor than would have occurred without the differential. In industrial terms this has been termed the 'brand effect', where a company with greater market share can use its advantage to compete more aggressively to capture even more of the market. The brand effect snowballed rapidly in this case because market share was the only difference. Five years into the simulation, the company which started with an edge in, say, Gulf Coast hurricanes held all the policies in that segment of the market.

Case 2: As for Case 1, except that there is a hurricane in one region in 2005

The next step was to examine what would happen if a disaster actually occurs. We took the Case 1 conditions as a starting point, then specified that a hurricane occurs in the Gulf Coast half way through the ten-year simulation, in 2005. This simulation showed that the brand effect could have a negative impact on a dominant market player, if nature deals it a joker. All companies started with equal capital. By the time the hurricane hit, the leading company had obtained all the hurricane insurance policies in the Gulf Coast, so had substantially increased its capital base. However, this was still less than its enormous exposure from its market dominance, so the company went bankrupt. After the disaster, the demand for hurricane insurance increased, causing the price of policies to go up. Companies which had managed to stay alive because of their smaller exposure could subsequently exploit this upturn in business. But the business recovery came too late for the bankrupt market leader. The hurricane didn't have a significant overall effect on either of the two active reinsurers. They survived the fall in reinsurance premiums in the first five years, so could benefit when property owners began to buy hurricane insurance again.

This outcome was produced by an early version of Insurance World in which bankrupt companies do not actually go out of business. Instead, they go into a kind of receivership known in the US as 'Chapter 11', where a company can stay in business after going bust. Future versions will allow for the disappearance of bankrupt companies, resulting in a redistribution of their customers to existing and new companies according to rules agreed by consortium members.

Case 3: As for Case 2, except that the Gulf Coast hurricane hits in 2001

Outcomes were very different when we ran a simulation similar to Case 2, but with the hurricane happening in 2001. In these circumstances, the primary insurer starting with the market-share advantage survived without going bankrupt. The hurricane occurred before this firm had acquired the whole of the Gulf Coast market, so it had to pay out much less than in Case 2 after the disaster occurred. All primary insurers shared the risk and payouts roughly equally because the extra market share held by the leading company in 2001 wasn't enough to make a big difference. When Gulf Coast property owners became sensitised to buying hurricane insurance after 2001, the market leader again exploited the brand effect to grab a bigger share of this business. It became by far the strongest company in Gulf Coast hurricanes, but this time it remained in profit and did not wipe out the market share of other companies. There was a dip in the capital for the two active reinsurers in 2002, but they recovered reasonably well for the rest of the simulation period.

Enhancing Insurance World to meet users' wishes

At each consortium meeting we have asked members for a 'wish list' of the things they would like to see in the next version. As many new wishes have been added between meetings as those addressed by the SFI model builders on the previous list. This is a natural part of this kind of collaborative, experimental development and an indication of the vitality of the project.

The kinds of wishes we have incorporated can be illustrated by some enhancements made for version 1.3, which was developed for a consortium meeting to be held shortly after this LSE seminar. For example, 1.3 satisfied the wish to enable bankrupt firms to be removed and their policies redistributed to other firms. We also introduced a number of new business features, like the capabilities to set limits on how much reinsurance business can be accepted and to allow for diversification by primary and reinsurance companies. Role playing was a particularly important capability introduced in 1.3. It lets people interact with the system to learn about the consequences of choosing different courses of action. We are also continuously improving the user interface to make it as appealing and efficient as possible for any employee in a member firm who might want to apply the model to their own business environment.

The most important planned enhancements relate to the built-in adaptation capability. Each primary insurer and reinsurer starts out by having access to their own basket of rules in the model. Adaptation will allow agents to monitor progress and change rules in the light of feedback from the unfolding simulation. Special adaptation rules will be needed to help agents determine under what circumstances a different rule should be moved to, and which rule would produce a better result.

We will also invoke a measure of innovation and creativity by having a procedure to enable an agent to create new rules which didn't exist before. Arthur and colleagues did this in their stock market world in an imaginative way. They encoded all possible trading rules as a 60-bit data stream, which has a capacity for 260 rules that could be translated into the everyday traders' language of buying, selling, or holding stock. Genetic algorithms and other procedures can then be applied to the data stream to generate different rules from those traders started with.

We are making good progress towards incorporating at least some primitive adaptation rules into Insurance World. One of the main reasons we did not include the crucial adaptation feature from the start is that people in the insurance industry do not yet have a good feel for the explicit rules or principles underpinning the decisions which they make all the time. This makes it difficult enough to define the basic set of agent rules. For adaptation, we would also have to add a whole bunch of even more complex meta-rules about how to evaluate and change existing rules.

Trying to identify rules when there is no agreed existing set is similar to eliciting knowledge for an expert system by trying to divine what people are thinking by looking at their actions. This is equivalent to mathematical 'inverse problems' that are hard to solve because there may be an infinite number of different rules which could generate the particular observed behaviour being analysed. For example, Ptolemy and Copernicus provided different rationales to explain the behaviour of the solar system, but it took much time and effort to decide which one offered the better understanding of what actually happens.

Learning how to make Insurance World more realistic

The catastrophe insurance industry is a very complicated system. Managers and decision-makers in it have to juggle many different aspects, with no single manager able to deal with all of them.

In order to get the initial version of Insurance World underway in the face of such variety and complexity, the model builders at SFI set default values for many basic parameters, such as those defining the profile for a particular insurer or reinsurer. Our lack of expertise in the insurance business meant we inadvertently created an imbalance between some default values and the magnitude of payouts after a disaster, resulting in some companies being driven towards bankruptcy too quickly after a relatively small number of disasters. The defaults are not hardwired, but take time to hand-tailor before a simulation if there is to be a more realistic trade-off between default capital values and the likely numbers of high-intensity events.

These are just a few of the vast number of variable in Insurance World which can be adjusted. We are learning all the time how to improve the balances between them. However, this kind of fine-tuning sparked a discussion at a recent consortium about why we didn't try to sort out the model's underlying rules and equations before we wrote its first line of code. Although that might seem to be a logical approach, good computer models are created by constructing the experimental artificial world in parallel with the equations that control it, in a similar way in which a book is written through a continuous process of revision and refinement, rather than by first trying to define an outline in painstaking detail in the expectation that it will hold good while the whole text is written.

Building an artificial world in a computer is, therefore, an intrinsically interactive, evolutionary process in which feedback from experiments and real-world knowledge inform the definition of the rules underlying the dynamics driving the system. This inevitably takes much time and effort if the world being created is as complex, and its rules as undefined, as in cases like Insurance World and Sainsbury's SimStore (Venables and Bilge 1998).

The value of Insurance World to its users

Business members of the Insurance World consortium come from a variety of backgrounds and joined the project for different reasons. Some are looking for a useful hands-on tool to help them make real decisions about their actual operations, such as in understanding what kinds of risks, at what levels, they should insure. A few companies joined primarily because it gave them an opportunity to meet other members regularly. They were not particularly looking for a tool to come out of the collaboration.

Other companies find the prime benefit results from their participation in the process of rethinking how their previously implicit assumptions about the way their industry operates can be made explicit as agent rules in the computer model. Delivery of a tool at the end of that process is seen by such consortium members mainly as useful icing on the cake. For example, all industry members at one meeting agreed the SFI modellers had set certain default values incorrectly. They then spent hours debating what values and equations would better reflect how things actually work. This process of learning by interacting with a model, and with each other, is helping consortium members to gain a deeper understanding of how their industry works.

There are also differences of opinion about the best use that could be made of any tool that is produced. Some firms want a tool for everyday use. Others want support for training their people. The role-playing capability will be of particular assistance for such training. It can help underwriters, managers and other new or existing employees to learn about the skills involved in their job, by seeing the way decisions they take affect Insurance World.

We promised to deliver a common-denominator 'plain vanilla' systems, incorporating all the basic ingredients all companies could feel comfortable, such as properties, events, and insurance and reinsurance companies. It would not represent any one company, but each consortium member could modify the vanilla model to reflect its own specific worldview, perhaps with the assistance of SFI. This approach fulfilled the SFI's charter, which prevents us from doing classified, proprietary, or contract research that cannot be published freely anywhere. We also believe it would have been unrealistic to try to build an initial model that sought to accommodate the different flavours requested by each firm.

The good progress we have made to date indicates that the collaborative, questioning and iterative nature of the Insurance World development will ensure the model we finally deliver will provide a realistic core simulation of the catastrophe insurance industry. The active involvement and constructive feedback from companies also makes us confident that consortium members will understand the model sufficiently well to enable them to tailor it to their own needs and to exploit its capabilities successfully.

References

Casti John L "Paradigms Lost" New York: Avon Books, 1990.

Casti John L "Complexification" New York: Harper Collins, 1994.

Casti John L "Would-be Worlds" Chichester: John Wiley & Sons, 1997.

Casti John L "The Cambridge Quintet: A Work of Scientific Speculation" London: Little, Brown, 1998.

Venables M and Bilge U "Complex Adaptive Modelling at J Sainsbury: The SimStore Supermarket Supply Chain Experiment" Warwick: ESRC Business Process Resource Centre, Warwick University, 1998.