papers on agent-based economics nr 5 Are Genetic Algorithms a good basis for economic learning models? Sylvie Geisendorf section environmental and innovation economics university of kassel papers on agent-based economics nr 5 Abstract: Genetic Algorithms (GA) have been used for some years now to depict learning in economic models. Some authors criticize their use on the basis that they are a biologically motivated procedure having nothing to do with human learning. One argument of this paper is that the criticism of GA is focused at the wrong point – and was probably incurred by the much too simple applications we have seen so far. It is not primarily the origin of a model we should be concerned of, but its general characteristics and the specification in its current use. After a brief introduction into the procedure the paper tries to show why GA offer some important features for the modelling of bounded rationality. Learning models based on them are among the few that create novelty and describe the mechanisms of selection, recombination and variation by which novelty is generated. The paper discusses the criticism of GA and argues that the biological origin of the model should not be a substantial problem. The biological features of the model are only a shell in which the general mechanisms of evolutionary processes are imbedded. An up to now underestimated problem however lies in the adequate specification of the fitness function of GA. Proponents as well as critics of GA seemed to have overlooked the necessary distinction between internal fitness criteria of the agents and external criteria of the economy. Both are relevant for economic selection processes and have their proper place in the model. If this is respected GA based learning models can be a useful tool to investigate economic evolution. Keywords: Evolutionary Economics, Genetic Algorithms, Learning, Bounded Rationality, Modelling, Methodological work . 2 papers on agent-based economics nr 5 I. Introduction Since Genetic Algorithms (GA) have been developed by John Holland (1975) as a model of biological evolution two things became obvious. First, the simulation model was a very good problem solving procedure, being able to detect almost optimal solutions even in complex environments. Hence a quite similar procedure, Evolutionary Strategies, had been developed by engineers to design technical components that had to satisfy complex requirements (Rechenberg 1973) and GA themselves have been used to solve logistic and technological problems (Goldberg, 1989). Second, and connected with this first observation, the mechanisms working in GA seemed to be common mechanisms of evolutionary processes in general. They are the mechanisms explaining the origin of novelty and why systems undergoing evolution are able to adapt so well to their environment. Both aspects seemed to qualify them as a means to describe economic evolution or learning and adaptation of economic agents. Unfortunately however they often seem to get mixed up. Economic learning models are told to be models of bounded rationality, but in quite a few models the algorithms are given all the information necessary for the detection of the optimum, enabling the agents to reach it after some time steps. A process isn’t necessarily a learning process, just because it reaches an optimal solution after some time or because new strategies are based on the evaluation and variation of former ones. Most models including GA so far have neglected this point. The procedure has been used to find out whether rational expectations can be learned (Lawrenz 2002). But it barely has been discussed on what kind of information such a “learning” process can be based. If you give GA all the relevant information about a system, it is able to optimize. But learning isn’t finding the right solution in a large heap of information. Learning is often restricted by a lack of information or a lack of abilities or motivation. And we still can use GA to illustrate what is going on when someone learns under all these restrictions. All we have to do is introduce them into the model. Unfortunately the specification of bounded rationality is not straightforward. There is only one kind of rational expectations, but a multitude of bounds to rationality. This puts rational expectation models at a seeming advantage, because once agreed upon this basic assumption, you only have to consider one type of agent and get a general result for all situations. The crucial point is, whether the results are correct, which they are often not as a considerable number of empirical studies show. 1 When depicting bounded rationality, the modeller has to decide carefully, which specific limitations on knowledge, competence and motivation are 1 See Conslik (1996) or Gintis (2000). 3 papers on agent-based economics nr 5 given in a particular case. The quality and applicability of the model depend on the quality of this input. Proponents of evolutionary models often apply them without giving much of a justification for their assumptions and opponents just criticize the approach in general, because it originates from biology without discussing critical details.2 Nevertheless the inclusion of bounded rationality and learning into economic models is important. We need theoretical insights into learning in economic contexts to understand how people, differing from neoclassical homo oeconomici, actually behave. Neoclassic economists often emphasize the importance of predictions. Optimization models, they argue, have the advantage of detecting precise solutions. But a solution is only as good as its correspondence to reality. If people are less well informed and less apt than the model assumes, they are not able to reach to unambiguous optimal solution. So the prediction is wrong. Consequently we have no other option than trying to model real people’s behaviour if we actually want to be able to predict what might happen, although prediction in that sense is no unique number but a probability distribution about possible outcomes. The paper argues that GA are a good basis for economic learning models, because they offer a frame including all aspects of the generation and diffusion of novelty in different kinds of systems. This frame has to be filled out accordingly – which admittedly often has been failed to do properly. Section 2 gives a brief introduction into the elements of GA. It explains why the mechanisms depicted by GA are so important for an endogenous model of the creation of novelty and gives some details about how they could be interpreted for economic learning models. Section 3 discusses whether models derived from biological principles of evolution are applicable to economic questions. Criticism directed against GA-features like pair-wise exchange is too superficial because such characteristics are easily changeable and not among the constitutive elements of the model. The paper sustains the opinion that GA are less biological than they might appear at first sight. They depict general principles of evolutionary processes, encased in a biological shell. All we have to do is strip them of this shell and fill the frame with economic content. There however the real difficulties start. Section 4 examines the up to now underestimated problem of how to design the fitness criteria of learning processes sensibly. When deciding about new solutions economic agents have to evaluate how well former solutions have performed. The success of these former solutions – or the suspected success of some variation or mixture of them – determines the basis on which new solutions will be developed. It is in this process of evaluation and selection that internal and external fitness criteria influence the decision process. The modeller has to distinguish them, 2 The most recent discussion on the transferability of Darwinian principles of evolution to economics is to be found in the Journal of Evolutionary Economics 16 (2006). 4 papers on agent-based economics nr 5 to design both appropriately and to put both at the right places in the model. In most GA based learning models it has been overlooked that were humans are concerned we have to distinguish between internal and external selection and that criteria of success held by the agents aren’t necessarily the ones of the external environment. This led to the above mentioned quite simple applications of GA, where agents were given a lot more information than they could reasonably be assumed to be able to procure themselves. To be quite clear about this point, GA based learning models are not the only ones confronted with this trouble. All models picturing the introduction or diffusion of novelty in the economic system have at least implicit fitness criteria. On the contrary it is one of the advantages of GA that they encompass an explicit fitness function, we are able to design accordingly. Other evolutionary models like replicator dynamics are structurally much too simple to do that at all. Section 5 concludes the paper and summarizes the main points. II. How Genetic Algorithms work GA consist of a number of binary strings, containing information about how to behave in their environment and some operators, changing the strings. After the instructions encoded in the strings have been carried out, they are evaluated by a fitness function, representing their environment. The better performing strings get higher scores. These in turn are important for the probability to be chosen by a selection operator, determining which strings are allowed to reproduce. The chosen strings then undergo a procedure of crossing-over and mutation, and the so built offspring forms next period’s generation that undergoes the same operations again and again. GA have been developed by John Holland (1975) as a model of biological or more precisely genetic evolution.3 The strings represent genes or chromosomes. Decoding them, means to build a phenotype that is acting in its environment. The requirements of the environment are represented by a fitness function that evaluates the success of the strings. Darwinian survival of the fittest is modelled by a probability to be chosen for reproduction, proportional to fitness. When two strings are allowed to reproduce, their genetic material is mixed by crossing-over, with a certain probability to keep the whole string of the original individuals. And there is a probability of arbitrarily switching a 0 to a 1 or vice versa, referring to mutations. This offspring can either replace their parents or be added to the population. In the 3 Introductions into GA besides Holland (1975) or Holland et al. (1986) are Goldberg (1989) or Mitchell (1997). 5 papers on agent-based economics nr 5 second case the least successful individuals are removed from the population to keep it constant. GA are a procedure that is very apt to find good or even optimal solutions to complex problems. For that reason they (or similar procedures) were also used to solve technical problems.4 And because they showed, how a population of searching agents could ameliorate its solutions over time, they started to be used as learning models in economics.5 When using GA in an economic context, the operators of the algorithm are interpreted as steps of a process of “adaptive learning” (Dawid, 1996). A string is an idea or a market strategy of an agent. Selection and the related fitness function depend on the agents’ success in their economic environment, and on the agents’ personal interpretation of the reason for this success.6 New ideas are developed by recombining or copying successful old ones (crossingover), including some mistakes or experiments (mutation). Although a string only contains zeros and ones it is able to encode almost everything, ranging in the economic domain e.g. from a simple number like production output to a complicated production scheme, including amounts of different input factors and e.g. production or marketing strategies. In the latter case the string is decoded sequentially, where each part of it stands for another characteristic. The decoded sequences can than undergo any number of additional transformations by equations or logical instructions to get the final effect upon its environment. Assume e.g. that a string encodes 3 input factors and the amount of money spend on advertisement. The input factors are fed into a production function, inferring an output and the corresponding production costs. Let’s further assume that sales depend on promotion effort. The benefit of a firm is thus determined by sold items minus production costs minus money spend on advertisement. Of course, the environmental side of the model is also open to any kind of complexity. To remain in the economic domain, we could assume that all products are sold for a fixed price, or we could e.g assume that prices are determined by demand and sales depend on first mover advantages or conformance of the advertisement strategy with preferences of the customers. It is this flexibility and the fact that GA are among the few models that do not only depict the diffusion side of the introduction of novelty, but 4 See Rechenberg (1973) or Koza (1992). 5 E.g. Holland/Miller (1991), Arifovic (1994 and 1999), Andreoni/Miller (1995), Dawid (1996 and 1996b), Dawid/Kopel (1998), Birchenhall (1997), Geisendorf (2001) and Lawrenz (2002). For a short description of these models see Geisendorf (2001). 6 However, the designer of the model has to decide, in what variable this success expresses itself: profits, market share, average profits of the last n periods, or other. We are coming back later to this crucial aspect. 6 papers on agent-based economics nr 5 also the generation side, that makes it worthwhile to consider their qualification as a model for learning and the development of novelty in economic contexts. III. Modelling learning by the mechanisms of biological reproduction? Originally, GA were designed to explain genetic evolution, but they do so in a very stylized way. As Burke et al. (1989) showed they do not seem to be much more sufficient from a biological point of view than from a psychological one.7 Saying this, the GAs ability to describe biological evolution should not be discredited. It should merely be illustrated that GA are a stylized procedure, capturing important details of evolution in general: selection, variation and retention (Hodgson 2002) that are all subject to some degree of randomness. Even authors, who pronounce themselves against the use of GA - or biological metaphors in general - as a model for economic evolution, are far from free of using them.8 The point is, that it seems to be quite difficult to imagine the generation of novelty without these operators.9 When novelty is generated it has to fulfil some requirements of its specific environment or of its designer who, in turn, has to consider the environment. It is judged in its ability to do so in comparison to other similar “things” (be they animals, people, artefacts or ideas) and its survival, use, price or other success depends on this ability. This is the selection side of the process or what happens to novelty after its generation.10 The important point here is not whether selection takes place intentionally or by some sort of external environment. The point is that it does take place and that is has consequences for the generation side of novelty. New things do not emerge out of the blue; they are made of material and ideas that already exited. In biology they are mixtures of their parents. Artefacts are made of new combinations of already existing materials, components and ideas. It is no coincidence e.g. that the first cars looked like coaches. But neither biology nor an inventor just takes some arbitrary base 7 “Putting more Genetics into Genetic Algorithms” (Burke et al. 1998). 8 One of the fiercest critics (Witt 1997), at other places, describes aspects of economic learning and development in exactly the terms of biologic evolution. He says, that ideas are generated by the recombination of known cognitive elements (1992: 5), “individual learning takes place in formally the same way as the genetic adaptation” (1992: 9) and competition is a selective process, in which the chance of survival depends on the surrounding variants (1992: 13). See also Metcalfe (1994: 29). 9 We can not go into much detail right here. This point has been discussed elsewhere (Geisendorf 2004). 10 A lot of models in Evolutionary Economics only consider this side, i.e. the diffusion of novelty, described by replicator dynamics (Metcalfe 1992, Witt 1992) or synergetics (Weise 1993, Weidlich/Braun 1992). 7 papers on agent-based economics nr 5 material. In both cases the choice depends on former success. In biology this is straightforward, because individuals who were to week to survive are simply not available for reproduction anymore and it is well-known that in a lot of species candidates fight over partners and the winner is allowed to reproduce. In the economy the same selection process takes place, although in a mixture of intentional internal selection and external selection by the market. The market selects products for their selling success, reflected in prices, sold quantities or market shares. This is the basis for the intentional selection by entrepreneurs or inventors, when designing new products. They know about the properties of existing products and have an idea about what kind of products customers would by and they consider these aspects when designing something new. Snowboards e.g. are a mixture of skies and skate- or surfboards, for the young and “cool”. They combine materials and technical properties from existing products and use the knowledge about a specific group of buyers for some of these products. External market selection gave them an idea about the possible success of a new product, mixing the properties of some formerly successful ones. Internal selection enters the development process, because people, unlike nature, are able to decide intentionally about the specific properties of the novelty, and are even able to perform tests before actually launching a new product. Whatever case we consider, there is some sort of selection and recombination and a little arbitrariness, due to mutations, mistakes or experiments. We will come back to the distinction between external and internal selection below. The main discussion on bounded rationality and learning is taking place in Evolutionary Economics. It was incorporated into models of norm building, technological development or asset buying behaviour and a considerable number of learning models were based on GA.11 Although it has been shown why the above mentioned mechanisms of selection, recombination and mistakes typical for GA are important when the generation of novelty should be depicted, there are some aspects of human learning, different from or missing in GA. The most evident criticism against the use of GA as a model for human learning is the fact, that it entails some typically biologic features like a pair-wise exchange of information, a 11 There are several different ways in which bounded rationality or learning has been tried to model. We will just mention some for reference. A good survey is Brenner (1999). Mathematical approaches are: Silverberg (1988), Weise (1993) who used a synergetic approach or Metcalfe (1994) with a model based on replicator dynamics. Models using Genetic Algorithm are: Holland/Miller (1991), Arifovic (1994), Andreoni/Miller (1995), Dawid (1996 and 1996b), Birchenhall et al. (1997), Dawid/Kopel (1998), Cooper (2000), Geisendorf (2001), Lawrenz (2002). 8 papers on agent-based economics nr 5 genotype/phenotype distinction and the lack of cognitive aspects of learning, like memory, internal selection or the motivation for change. Human information exchange does not necessarily take place pair-wise. Though it is not obvious to decide ad hoc what kind of information exchange would be more adequate. Some kinds of information, such as statistical data, the stock market development or TV news, are available to (almost) everybody.12 Note however that most people do not procure themselves all the information potentially available to them and/or are not necessarily able to process it in a sensible way. Other data are only attainable through direct relationships between two or more people, but might be transmitted in one direction only. The specification of the learning algorithm therefore requires a decision about the adequate kind of information transmission. Technically, an alteration of the procedure is no problem at all. A second difference between biological and human information transmission is the genotype/phenotype-distinction in biology. Translated to human learning, GA imply that only the phenotype is visible, but it’s still the genotype, that is transmitted. According to Chattoe & Gilbert (1998) it would be more plausible to assume telepathy straight away, than to believe, that entrepreneurs are able to derive other people’s strategies from their market results. It is obvious, that just having a look at a firm’s profit does not allow a reconstruction of its production scheme. But that, of course, is not what happens. An agent’s success only transmits a first orientation for its competitors. It is a selection criterion. Once selected, the competitors try to find out the reasons for the success. It is true, that GA do not model the process of information gathering, but we can assume that it is taking place in the background. Beckenbach (1999) took that point into account by associating crossing-over activity with search costs. More important are aspects of learning that are missing in GA. GA are an aimless searching procedure. There is no conscious intention and no reflection in its search, and there is no awareness of past experiences. The most undisputed point may be the lack of a memory. It is true, that a memory is only advantageous if similar situations are recurring and that “In nonstationary environments, agents are always bound to try to understand future behaviours of other agents or future events that had never occurred in the past“ (Dosi & Egidi 1991: 151). But still, we can assume that at least some patterns are reappearing and that experience plays an important role in human decisions. In Geisendorf (2001) it was tested if the inclusion of a memory into the algorithm makes a difference in some fishery models. Surprisingly it did not. 12 This kind of information availability is implicitly assumed by synergetic models in social sciences. Everybody knows what everybody else is doing - or more precisely, what percentage of the population is doing one thing or another (see Weidlich/Braun 1992 or Weise 1993). 9 papers on agent-based economics nr 5 In the profit oriented case, the only information the agents could obtain were catching effort, landings and corresponding profits. But profits are not exclusively related to catching effort. Although not altering the results in a considerable way, the inclusion of a memory was a worthwhile move, because it shed some light on an interesting observation: in complex decision contexts, we are often confronted with uncertainty or ignorance, and a good knowledge and remembrance of economic facts does not remedy the lack of other, e.g. ecological, information. The market punishes wrong behaviour, but it is not able to provide the agents with all the needed knowledge, so they are bound to make the same mistakes over and over again. Remembrance is closely related to another cognitive aspect: internal selection. If the agents have a memory, they can compare new strategies with former ones or verify, whether a “new” idea has already been tried and didn’t work. Even without a memory, after generating a new strategy, an agent could have a break and think about its potential, given last periods information, instead of just executing it blindly. This amounts to internal selection. Some authors even argue that external selection does not play an important role in economic contexts at all (see Witt 1997 or Brenner 1998). “Fitness in social evolution is rather subjectively determined” (Brenner 1998: 19). Inasmuch as it’s certainly true, that subjective fitness criteria play an important role in human decisions, it is unrealistic that external restrictions are irrelevant when it comes to market survival. In some economic applications of GA internal selection is included by an “election operator” (Arifovic 1994). New strategies are only put into practice if they would have been more successful under last period’s circumstances.13 This suggestion can be considered when specifying the model. But it might be a too strict avoidance of experiments and error making. And it is no good idea for situations in which the agents know the economic environment is changing and former conditions are not translatable to future success. In such cases the formulation of a more complex mental model by which the agents store their inner representation of the external world is required. In any case, a strategy that was matched against others, using some internal model, and selected for actual implementation, would than compete against other agent’s internally selected strategies on the external market. We will come back to this important distinction in the following subsection. 13 Another way to include internal selection is Holland´s Classifier System (Holland et al. 1986 or Holland & Miller 1991) or the use of a second, internal GA as a selection mechanism between ideas, where all known strategies are measured against a subjective fitness function, proposed by Beckenbach (1999) or Chattoe & Gilbert (1998). 10 papers on agent-based economics nr 5 The last missing cognitive element in GA to be discussed is the lack of intentional change (see Brenner 1998). Strategies are chosen and modified all the time, whereas real people need a motivation for change. Simon (1997) states, that people are only looking for better solutions, when dissatisfied with their current results. In some mathematical models, an aspiration level, that needs to be unfulfilled to incite action, is defined as the ambition to be at least as good as the average (e.g. Ebeling et al. 1997). The inclusion of this condition in the above mentioned fishery model slowed down the rate of change and stabilized the outcomes, without however, altering the general quality of the results in most cases. Only in some specific constellations, the stabilization prevented overexploitation, because everybody was so content with his income, that no growing of the economy took place (Geisendorf 2001). Apart from the effects on the results, we should be aware of what we are doing, when including an explicit motivation level to be surpassed, before changes could occur. Such an operator implies that well performing agents do not learn and only less successful ones are allowed to improve. Whereas it is certainly true, that people do not alter their strategies all the time, and the relation between strategy improvement and fitness is less straightforward than in biology (Chattoe & Gilbert 1998), it still appears more plausible that - analogous to biologic evolution - the more successful individuals are more likely to have the financial and intellectual potential for further ameliorations. The above paragraphs should have shown the value of a procedure like GA to encompass the whole process of the endogenous generation and diffusion of novelty in a system and they should demonstrate that the algorithms provide a mere shell or basis for the design of learning models where all the aspects deemed important can be included.14 It ought not primarily to be the origin of a model we have to worry about. What is worrisome is the often quite careless application of GA in economic contexts. It is not the model in itself that is often inappropriate, but its explicit specification. IV. How to configure fitness criteria for learning algorithms The above issue has not only been discussed in some detail because it still represents the most prominent criticism against a certain way of modelling, but as well, because it already points to the more interesting general problem of the adequate design of learning models. When investigating the justification of the above criticism, it should become visible that a 14 Of course there are models of biologic origin, like Replicator Dynamics, that do not offer much room for specification and have to be rejected altogether for most contexts (Joosten 2006). 11 papers on agent-based economics nr 5 satisfactory specification of economic learning models is far from obvious. Interesting enough however, this problem has not been discussed thoroughly by the authors using or disapproving of such models. An exemption is Beckenbach (2005) who proposes to base the design of economic learning models on insights of cognitive psychology, albeit on a theoretical basis. Until what point are we able to capture human learning? In GA, as well as in other simulations or mathematical learning models, this question is closely related to the fitness criteria. Mathematical models do not necessarily have a fitness function, but they still do have some fitness criteria, objectives or preset behavioral rules, that are all subject to the same problems we are going to discuss now. It is, by the way, one of the advantages of GA that they have an explicit fitness function we are able to design suitably. The algorithm is not exactly an optimization technique, but it has the ability to adapt well to given problem structures and produces nearly optimal results for a lot of cases. When using such a model to depict human learning, this is the first point a scientist has to be aware of – but often does not seem to be. Using such a procedure to answer the question whether optimal behaviour can be learned (Lawrenz 2002) is sort of a tautology. Instead of calculating the optimum analytically, another method, able to approximate optimal solutions, is taken. This entails the advantage that the stepwise approach and improvement of the solution makes the existence of an underlying learning process more plausible. It probably is tempting to use GA to prove, that it is possible to reach the optimum by some kind of consecutive progression. The necessary assumptions though, that would allow for such an advance, are sometimes not much more realistic than assuming perfect rationality straight away. Such models have either been written to prove the possibility of optimization or it was overlooked that the algorithm has been endowed with the ability to optimize. And this is true for other learning models as well. Baysian updating of expectation is a good example, as are the least squares method, gradient climbing or replicator dynamics. Two things have to be pointed out here. First, it should be evident that the possibility to describe the improvement of economic strategies by a stepwise procedure is no evidence for the possibility of learning. It might just be the exchange of one optimization technique against another, possibly less perfect one, which makes it look as if the restrictions to optimization had been taken into account. But a deviation from the optimal solution does not make the procedure more realistic per se. Whatever outcome a model generates, the real restrictions to optimization usually have not been taken into account and this is a shortcoming of most economic learning models. This leads to the second point. The above criticism does not necessarily mean GA based learning models are of no use. It implies that there is no obvious 12 papers on agent-based economics nr 5 and unique way to do set them up. GA offer a considerable openness concerning the specification of the agent’s abilities and the adequacy of the model depends on the assumptions we put into it. Unfortunately, in a lot of GA based learning models, the agents are enabled to find the optimal solution, even though bounded rationality should be depicted. Let us now illustrate why. The fitness function is probably the most important part of GA. In biology, fitness is an external criterion. Nature “decides” who is going to survive and reproduce, without the necessity for the individual to know about natures criterions. There is no conscious attempt to change strategies in order to be selected for survival. In the economy something similar exists, because in the end, market decides, whether an economic agent survives, but in-between the individual agent might very well pursue quite different aims. He might believe e.g., that large profits will guaranty survival, whereas the market decides for the survival of someone, who accumulated market shares at the cost of low initial profits, but in the long run could outperform all competitors and establish a monopolistic position. In Geisendorf (2001) two groups of fishermen have been compared, one who was profit seeking, but ignoring ecological constraints, and one who was trying to maintain fish stocks within sustainable bounds. The profit oriented group was doing much worse from an economic point of view, than the one who was ecologically aware. The later were making much more money, which they might or might not care about. At least, making money was not included in their fitness function. We could just as well assume, they were only interested in the environment, than that they knew maximum sustainable yield might be a good strategy to make high long term profits. But if they knew, they should have tested this hypothesis against reality, i.e. we would have to include it in their internal fitness criteria. It is important to distinguish carefully between intentions of the agents and demands of the market. On one hand there are aims of economic agents, their internal models about how to reach their objectives and available information they could use to build mental models and to evaluate success. This is what we have to put into an algorithm when designing the learning process. On the other hand there is reality. Market exigencies can deviate in two ways from individual aims. Agents might try to do the right thing, but be unable to accomplish it, due to wrong internal models or a lack of information or the goals they pursue might well be reached but not be fit for market survival. In both cases, market exigencies must not be part of the fitness function of GA. They have to be included in the form of external restrictions, where “external” means, external to the learning process but not to the dynamics of the model. Economic requirements influence the behaviour of agents, without being part of the deliberate decision process. Of course, agents try to get a good representation of the real word, but it is 13 papers on agent-based economics nr 5 often explicitly the gap between internal and external fitness criteria, i.e. between internal mental model and real world dynamics that forms the restriction we try to depict when modelling bounded rationality or learning. The consequences of this distinction between individual and market requirements go beyond having to put the right restrictions at the right place of the model. It is not only a question of deciding, what has to go into the fitness function and what must be part of additional constraints. The problem underlying this divergence is the fact that, whereas there is only one type of perfect rationality, there is a multitude of bounds to rational decisions. A (representative) agent in neoclassical optimization theory is easy to model, because the basic assumption of rationality, i.e. conformity to market requirements, is fixed. When we want to model bounded rationality, there are lots of different ways of doing so, and neither the observable market results, nor the ideal market equilibrium, are the only - or even the most reliable - indicator of what people are actually trying to learn. How are we to know what information and internal models decision makers are actually using and what intentions they have? From a theoretical point of view it is interesting to experiment with different hypothesis and investigate the effects of various assumptions about the information basis and the decision processes. Simulation models are a very useful tool for this kind of research. As in all agent based computational models it is a good idea to put different assumptions into the model and compare the impact they have on the results (Pyka/Fagiolo 2005). Some results e.g. might be robust against a large span of assumptions. Where that’s not the case we know that we have to be particularly careful with our hypotheses. Especially in such sensitive cases a better theoretical foundation of the learning process, like proposed by Beckenbach (2005), and empirical surveys are important. V. Summary Since Genetic Algorithms (GA) have been developed by John Holland (1975) as a model of biological evolution, it became obvious that the procedure illustrates quite general mechanisms of the generation and propagation of novelty in different systems. GA consist of a number of binary strings, containing information about how to behave in their environment and selection, recombination and mutation operators, changing the strings. They depict the evaluation by which existing solutions to a problem are assessed in comparison to other solutions. Better performing solutions are selected as a basis for further ameliorations, which are built by recombining successful variants and making arbitrary changes due to mistakes or 14 papers on agent-based economics nr 5 experiments. As these steps are a good description of innovation or learning processes as well, GA have soon entered the economic domain and been used to depict learning in Evolutionary Economics. A string is an idea, a product or a market strategy of an agent. Selection and the related fitness function depend on the agents’ success in their economic environment, and on the agents’ personal interpretation of the reason for this success. New ideas are developed by recombining or copying successful old ones, including some mistakes or experiments. Some authors criticize the application of GA as a learning model on the basis that they illustrate a biological process having nothing to do with human intentional behavior. The paper discussed the objections point by point and showed that they are less relevant than it might appear at first sight. Typically biological features of the original GA, like pair-wise exchange or a purely external selection are no substantial characteristics of the model. They are due to their biological origin but can easily be changed and adapted to the process of human learning. An up to now underestimated problem however lies in the adequate specification of the fitness function underlying the selection process of the model. Proponents as well as critics of GA seemed to have overlooked the necessary distinction between internal fitness criteria of the agents and external criteria of the economy. Some critics state that fitness in economic evolution is mainly subjectively determined (Witt, 1997, Brenner, 1998). The question however is not whether economic selection is internally or externally motivated. In most cases both are likely to be relevant. The external economic environment evaluates products by attributing them prices, market shares, sold quantities or order figures. Individual internal fitness criteria are based in some way upon such an external selection and can contain other criteria like expected success of a new product based on market research. Both selection levels have to be included into the model at their appropriate place. The main problem of economic learning models based on GA so far is that they get mixed up. The agents are mostly endowed with information from the external selection process, they are unlikely to be able to procure themselves in reality, and their subjective fitness criteria are disregarded. Subjective fitness criteria however often deviate from market exigencies. They reflect mental models of the agents about how the world works, which are not necessarily right. If we introduce them into the learning algorithm, the model is allowed to make the same mistakes as human beings and starts to become a good model. Like in all models we have to make adequate assumptions. 15 papers on agent-based economics nr 5 References Andreoni, J. and Miller, J. (1995): Auctions with Artificial Adaptive Agents. In: Games and Economic Behaviour. 10: 39-64 Arifovic, J. (1994): Genetic algorithm learning and the cobweb model. In: Journal of Economics and Control, 18: 3-28 Arifovic, J. (1999): Inflationary deficit financing in an Open Economy: Evolutionary Dynamics. Santa Fe Working Paper 99-05-038E Beckenbach, F. (1999): Learning by genetic Algorithms in Economics? In Brenner T. (ed.) (1999): Computational Techniques for Modelling Learning in Economics. Advances in Computational Economics 11, Dordrecht: Kluwer Academic Publishers: 73-100 Beckenbach, F. (2005): Knowledge Representation and Search Processes – a Contribution to the Microeconomics of Invention and Innovation. Volkswirtschaftliche Diskussionsbeiträge Nr. 75/05, Universität Kassel Birchenhall, C. Kastrinos, N. and Metcalfe, S. (1997): Genetic algorithms in evolutionary modelling. In : Evolutionary Economics 7: 375-393 Brenner, T. (1998): Can evolutionary algorithms describe learning processes? In: Evolutionary Economics 8: 271-283 Brenner, T. (1999): Modelling Learning in Economics. Edward Elgar, Cheltenham, Brookfield Burke, D. et al. (1998): Putting More Genetics into Genetic Algorithms. In: Evolutionary Computation Vol. 6/ Issue 4: 387-410 Chattoe, E. and Gilbert, N. (1998): Just How (Un)realistic are Evolutionary Algorithms as Representations of Social Processes? Unpublished working paper for the workshop ”Agentbased and Population-based learning in Economics” 2.-3. 3. 1998 in Jena Conslik, J. (1996): Why Bounded Rationality? In: Journal of Economic Literature, June 1996: 669-700 Dawid, H. (1996): Adaptive Learning by Genetic Algorithms: Analytical Results and Applications to Economic Models. Springer Dawid, H. (1996b): Learning of cycles and sunspot equilibria by Genetic Algorithms. In: Evolutionary Economics 6: 361-373 Dawid, H. and Kopel, M. (1998): On economic applications of the genetic algorithm: a model of the cobweb type. In: Evolutionary Economics 8: 297-315 Dosi, G. and Egidi, M. (1991): Substantive and procedural uncertainty. In: Journal of Evolutionary Economics, 1: 145-168 16 papers on agent-based economics nr 5 Ebeling, W. Scharnhorst, A. and Karmeshu (1997): Economical and Technological Search Processes in a Complex Adaptive Landscape. Proceedings of the Workshop on Econophysics, 21-27 July 1997 in Budapest Geisendorf, Sylvie (2004): Der ökonomische Evolutionsbegriff: Selbstorganisation und Selektion. Lehmann-Waffenschmidt, M. Ebner, A. Forndahl, D. (eds.) (2004): Institutioneller Wandel, Marktprozesse und dynamische Wirtschaftspolitik. Perspektiven der Evolutorischen Ökonomik. Marburg, Metropolis: 79-96 Gintis, H. (2000): Beyond Homo oeconomicus: evidence from experimental economics. In: Ecological Economics 35: 3 Special Issue: the Human Actor in Ecological-Economic Models. p. 311-322 Goldberg, D.E. (1989): Genetic Algorithms in search, Optimization and Machine Learning. Addison-Wesley, Reading, Bonn u.a. Hampicke, U. (1992): Ökologische Ökonomie. Westdeutscher Verlag Hodgson, Geoffrey M. (2002): Darwinism in economics: from analogy to ontology. In: Journal of Evolutionary Economics (2002) 12: 259-281 Holland, J.H. (1975): Adaptation in natural and artificial systems. University of Michigan Press Holland, J.H., Holyoak, K.J., Nisbett, R.E. and Thagard, P.R. (1986): Induction. MIT Press Holland, J.H. and Miller, J.H. (1991): Artificial adaptive agents in economic theory. In: The American Economic Review 81: 365-370 Janssen, M. and Vries B. de (1998): The battle of perspectives: a multi-agent model with adaptive responses to climate change. In: Ecological Economics, 26: 43-65 Joosten, R. (2006): Walras and Darwin: an odd couple? In: Journal of Evolutionary Economics, 191: 561-573 Koza, J.R. (1992): Genetic Programming. MIT Press Lawrenz, C. (2002): Can Evolutionary Processes Result in Rational Expectations? In: Lehmann-Waffenschmidt, M. (eds.) (2002): Perspektiven des Wandels: Evolutorische Ökonomik in der Anwendung. Metropolis: 519-546 Metcalfe, J.S. (1992): Variety, Structure and Change. In: An Evolutionary Perspective on the Competitive Process. Papers on Economics and Evolution #9201. European Study Group for Evolutionary Economics Metcalfe, J.S. (1994): Competition, Fisher’s principle and increasing returns in the selection process. In: Journal of Evolutionary Economics, 4: 327-346 Mitchell, M. (1997): An Introduction to Genetic Algorithms. 3. Edition, MIT Press, Cambridge u.a. 17 papers on agent-based economics nr 5 Moxnes, E. (1998): Overexploitation of renewable resources: The role of misperceptions. In: Journal of Economic Behavior and Organiszation, 37: 107-127 Pyka A. and Fagiolo, G. (2005): Agent-Based Modelling: A Methodology for NeoSchumpeterian Economics. Volkswirtschaftliche Diskussionsreihe, Nr. 272, Institut für Volkswirtschaftslehre, Universität Augsburg Rechenberg (1973): Evolutionsstrategie. fromann-holzboog Verlag Silverberg, G. (1988): Modelling economic dynamics and technical change: mathematical approaches to self-organisation and evolution. In: Dosi, G., Freeman, C., Nelson, R., Silverberg, G. and Soete, L. (eds.) (1988): Technical Change and Economic Theory. Pinter, London Simon, H.A. (1997): Models of bounded rationality. Vol. 3: Empirically Grounded Economic Reason. MIT Press, Cambridge, Massachusetts, London Weidlich, W. and Braun, M. (1992): The master equation approach to nonlinear economics. In: Evolutionary Economics, 2: 233-265 Weise, P. (1993): Eine dynamische Analyse von Konsumtionseffekten. In: Jahrbuch für Nationalökonomie und Statistik, 211: 159-172 Witt, U. (1992): Evolutionary Economics - an Interpretative Survey. Papers on Economics and Evolution #9104 Universität Freiburg, revised Version Witt, U. (1987): Individualistische Grundlagen der Evolutorischen Ökonomik. Siebeck Witt, U. (1997): Economics and Darwinism. Papers on Economics and Evolution #9705. Max-Planck-Institute for Research into Economic Systems, Jena Impressum: papers on agent-based economics Herausgeber: Universität Kassel Fachbereich Wirtschaftswissenschaften (Prof. Dr. Frank Beckenbach) Fachgebiet Umwelt- und Innovationsökonomik Nora-Platiel- Str. 4 34127 Kassel www.ivwl.uni-kassel.de/beckenbach/ ISSN: 1864-5585 18
© Copyright 2025 Paperzz