Sunday, January 17, 2016

Growing Artificial Societies

_Growing Artificial Societies:  Social Science from the Bottom Up_ by Joshua Epstein and Robert Axtell  Brookings Institution Press/MIT Press, 1996  ISBN 0-262-55025-3

(1)  "Another fundamental concern of most social scientists is that the rational actor - a perfectly informed individual with infinite computign capacity and who maximized a fixed (nonevolving) exogenous utility function - bears little relation to a human being."
(7)  "One theme that runs through this entire book is _social connection_."
(10 -11)  "In particular, we find that the carrying capacity of the environment is _increased_ when agents trade.  However, this salutary result does not come free, for under some circumstances trade increases societal inequality."
(11-12)  When  agents have finite lives and the ability to change their preferences - "Nothing like the equilibrium of neoclassical theory emerges.  Of course, the laissez-faire argument is precisely that markets, left to their own devices, allocate goods and services efficiently.  the theoretical ccase for this is the so-called First Theorem of welfare economices,  However, when markets fail to arrive at equilibrium, the First Welfare theorem does not apply and this case for laisez-faire is undermined."
(19)  "The broad aim of this research is _to begin the development of a more unified social science, one that embeds evolutionary processes in a computational environment that simulates demographics, the transmission of culture, conflict, economics, disease, the emergence of groups, and agent coadaptation with an environment, all from the bottom up_.  Artificial society-type models may change the way we think about _explanation_ in the social sciences."
(20)  "Perhaps one dya people will interpret the question, 'Can you explain it?' as asking, 'Can you grow it?'"
(21)  ""Event unfold on a 'sugarscape.'  This is simply a spatial ditribution, or topography, of 'sugar,' a generalized resource that agents must eat to survive.  The space is a two-dimensional coordinate grid or lattice.  At every point (x,y) on the lattice, there is both a sugar level and a sugar capacity, the capacity beign the maximum value the sugar level can take at that point."
(22)  "The sugarscape wraps around from right to left (that is, were you to walk off the screen to the right, you would reappear at the left) and from top to bottom forming a doughnut - techically, a torus."
(23)  "Each agent is characteruzed by a set of fixed and variable states.  For a particular agent, its genetic characteristics are fixed for life while its wealth, for instance, willv ary over time."
"One state of eahc agent is its location ont he sugarscape... One might thing of an agent's initial position as its 'environmental endowment.'"
(23-24)  "Eacha gent had a genetic endowment consisting of a sugar metabolism and a level of vision.  Agents have different values for these genetic attributes;  thus the agent population is _heterogeneous_."
(23)  "All agents are given some initial endowment of sugar, which they carry with them as they move about the sugarscape.  Sugar colected but not eaten - what an agent gathers beyond its metabolism - is added to the agent's holdings.  There is no limit to how much sugar an individual agent may accumulate."
(25)  "The agents are also given a movement rule."
NB:  What if the mirror world is not a real-time updatable model but a simulation matched and tracked by actual data?  Or both?
(26)  "All the ingredients are now in hand.  We have a sugarscape and an initial population of agents, each of whom comes into the world with an environmental and genetic endowment, and we have simple behavioral rules for the sugarscape and the agents.  Initially there will be only one rule for the agents and one for the sugarscape, but in subsequent chapters both the environment and the agents will execute multiple rules."
(33)  "While initially quite symmetrical, the distribution ends up highly skewed.*  Such skewed wealth distributions are produced for wide ranges of agent and environment specifications.  They seem to be characteristic of heterogeneous agents extracting resources from a landscape of fixed capacity.  By contrast, the distribution of _income_, defined as the amount harvested per period less metabolism, is much less skewed.

*"Agents having wealth above the mean frequently have both high vison and low metabolism.  In order to become one of the very wealthiest agents one must also be born high on the sugarscape and live a long life."

"This distribution is our first example of a so-called _emergent structure_."
(35)  "We use the term 'emergent' to denote _stable macroscopic patterns arising from the local interaction of agents_."
"Understanding how simple local rules give rise to collective structure is a central goal of the sciences of complexity."
(36-37)  "The nature of the Gini coefficient or ratio is conveniently explained by reference to the so-called Lorenz curve.  This is a plot of the fraction of total social wealth (or income) owned by a given poorest fraction of the population.  Any unequal distribution of wealth produces a Lorenz curve that lies below the 45 degree line - the poorest X percent of the population controls less than X percent of the society's total wealth.  The Gini ratio is a measure of how much the Lorenz curve departs from the 45 degree line.  If everyone has the same amount of wealth the Gini ratio is zero, while if a single individual owns everything then the Gini ratio is one.  As the Gini coefficient increases society becomes less egalitarian."
(38-39)  ".., the so-caled Van Nreumann enighborhood, defined to be the set of sites immediately to the north, south, east, and west of a particular site....  And alternative is the Moore neighborhood, which includes all four sites of the von Neumann neighborhood as well as the four sites along the diagonals."
(42)  "While these waves seem to quality as emergent structure, the diagonal _direction_ in which they propagate is perhaps even more interesting.  Recall that during a single application of M the _individual_ agent can only move north, south, east, or west.  Yet the collective wave is clearly moving northeast - a heading unavailable to individuals!  On closer examination, the collective northeast direction results from a complete interweaving of agents, none of whom can move in this direction...  Here, the local rule _precludes_ individual behavior mimicking the collective behavior." 
*"In pure cellular automata (CA) models, waves are phenomena of significant interest.  Recently Sato and Iwasa ahve produced these in a CA model of forest ecology. REcent attempts by mathematical biologists to model the wavelike movement of certain mamal herds include Gueron and Levin and Gueron, Levin and Rubenstein.  For an economic model of 'herding' see Kirman."
(48)  "Pollution diffusion rule D:
Each eta time periods and at each site, compute the pollution flux - the average pollutio level over all von Neumann neighboring sites;
Each site's flux becomes its new pollution level.

"The reader with a knowledge of cellular automata (CA) will notice that this rule, which relates the pollution on any site to that on other sites makes the sugarscape a true CA."
(50)  "Subsequently, when diffusion is turned on, the pollution quickly spreads more or less uniformly around the landscape and many agents move back to the regions of highest sugar.  As they continue to gather and metabolize sugar, pollution increases while diffusing over the entire landscape.  There is a kind of rising 'red tide' that diminishes the welfare of all agents still alive on the sugarscape.*

"It turns out that the same type of dynamic pattern appears when the externally involved is positive rather than negative.  Positive externalities - increasing returns or network externalities, for example - give agent reasons to associate with one another, to spatially cluster.  Of course, the two types of externalities can be combined: there may be positive externalities associated with production but negative externalities associated with consumption.  Are cities the 'balance points' between these opposed effects?"
*"It is at this point that a clear 'tragedy of the commons' interpretation of  life on the sugarscape manifests intself.  Metabolisms are constant and so, for a fixed agent population, sugar consumption and the pollution it generates are fixed;  thus, the only way pollution levels can be decreased is to reduce the amount of pollution generated though production (harvesting) activities. if, for instance, those agents who harvest more sugar than they consume in following M were to follow some alternative rule, harvesting only, say, half the sugar they find beyond their metabolic needs, then overall pollution levels would fall.  While this behavioral rule would make these agents worse off, in comparison to M, by lowering their income, perhaps all agents could be made better off through side payments.  Alternative rules - institutions - for managing such common property resource problems in general are investigated at length by Ostrom and Ostrom, Gardner, and Walker."
(51)  "In Chapter III, we introduce combat.  Its intensity can grow when competition for resources becomes severe.  An influx of environmental refugees suddenly  boosts the agent density in the receiving zone and, naturally, competition for sugar intensifies dramatically.  The model suggests, therefore, that _environmental degradation can have serious security implications_."

"Paraphrasing, it amounts to the instruction:  'Look around for the best free site;  go there and harvest the sugar.'  And yet, all sorts of unexpected things emerge from the _interaction_ of these agents:  basic principles like the existence of environmental structures like waves that move in directions unavailable to individuals;  and biological processes like hibernation and migration (refugees).  And that strikes us as surprising.  The nature of the surprise is worth discussing."
(52)  "In short, _it is not the emergent macroscopic object per se that is surprsing, but the generative sufficiency of the simple local rules_."
(52-53)  "The Sugarscape model can function as a kind of _laboratory_ where we 'grow' fundamental social structures _in silico_, thereby learning which micromechanisms are sufficient to generate macrostructures of interest.  Such experiments can lead to hypotheses of societal concern that may subsequently be tested statistically against data."
(63)  "_Sustainable coevolution with one's environment_ is a necessary condition for "fitness,' if we with to retain this term at all."
(68)  "Interestingly, some 'Social Darwinists' oppose wealth transfers _to the poor_ on the ground that the undiluted operation of selective pressures is 'best for the species.'  Conveniently, they fail to mention that intergenerational transfers of wealth _from the rich to their offspring_ dilute those very pressures."
(80-82)  "Networks such as those we have described manifest themselves in the real world in many important ways.  Politically, restrictions on freedom of assembly, freedom of speech (press censorship), and freedom of movement (internal passport requirements) are standard tactics of repressive governments.  The main aim of these measures is to keep individual dissenters - of which there may be a great many - isolated from one another, to keep them from _connecting_ with other dissenters, and so to thwart the emergence of an organized _community_ of dissenters, conscious of their numbers.  how do changing political borders and 'information revolutions' (for example, the Internet) affect the emergence of groups?  How does 'samizdata spread across a landscape?  Artificial societies allow us to study such questions systematically."
(86- 90)  "To complexity scientists, the moral is clear:  When you change local rules you may change emergent collective structures.  For policymakers, there is a corollary:  The most effetive way to alter collective patterns of behavior may be from the bottom upl by modifying local rules."
(95)  "The main issue we address is the extent to which interacting agents are capable of producing _socially optimal_ outcomes, that is, allocations of resources having the property that no agent can be made better off through further trade.  The artificial societies modeling approach allows us to explore such questions systematically and reproducibly.  In particular, we compare the performance of distinct classes of agents - neoclassical agents and various non-classical ones.  We find that neoclassical agents trading bilaterally are able to approach, over time, a price close to that associated with an optimal allocation.  However, when the agents are made progressively less neoclassical - when they are permitted to sexually reproduce or have culturally varying preferences - the markets that emerge generally have suboptimal performance for indefinite periods of time.

"Such results have important implications.  First and foremost, the putative case for laissez-faire economic policies is that, left to their own devices, market processes yield equilibrium prices.  Individual (decentralized) utility maximization at these prices then induces Pareto optimal allocations of goods and services.  But if no price equilibrium occurs, then the efficiency of the allocations achieved becomes an open question and the theoretical case for pure market solutions is weakened.

"We also investigate the effect of trade on variables studied in previous chapters.  We find that the _carrying capacity_ of the resource-scape is increased by trade, but so is the _skewness of the wealth distribution_.  More agents exist in a society that engages in trade, but the resulting society is more unequal.  Furthermore, the markets that result from our local trade rule generate _horizontal inequality_ - agents with identical endowments and preferences end up in different welfare states. Importantly, the welfare theorems of neoclassical economics do not hold in such markets."
(99)  footnote 8 "Below we show that the genreal effect of trade is indeed to _augment_ the carrying capacity."
(101-102)  "Agents move around the resource-scape following M, but are now permitted to trade with the agents they land next to, that is, their von Neumann neighbors.  When an agent-neighbor pair interacts to trade, the process begins by having each agent compute its internal valuations of sugar and space.  Then a bargaining process is conducted and a price is agreed to.  Finally an exchange of goods between agents occurs if both agents are made better off by the exchange.  The process is repated until no further gains from trade are possible."  
(102)  footnote 11 "... it is a kind of Edgeworth barter process."
(108)  "The immediate question for us - having banished the auctioneer and all other types of nonlocal information - is _whether our population of spatially distributed neoclassical agents can produce anything like an equilibrium price through local interactions alone.  It turns out that there is a definite sense in which they can!  However, the character of the equilibrium achieved by our agents is rather different from that of general equilibrium theory, for the markets which result produce less than optimal agent welfare - the potential gains from trade are not fully extracted - despite essential convergence to the general equilibrium price.  Furthermore, when we relax certain neoclassical assumptions (infinitely lived agents, fixed preferences) overall market performance is further degraded."
NB:  Does that mean working on neoclassical theories in the real world results in a falsely degraded market?
(113)  "This brings us to the so-called First Welfare Theorem of neoclassical economics.  This result is the foundation for economists' claims that markets allocate goods to their optimal social uses.  The theorem states that Walrasian equilibria are Pareto-efficient.  They are states in which _no realllocation exists such that an agent can be made better off without making at least one other agent worse off_.  But in statistical equilibrium

the First Welfare Theorem should be revised to say that a market equilibrium approximates but cannot achieve a Pareto-efficient allocation.  How close a given market comes to Pareto-efficiency
can be measured by the price dispersion in transactions. [Foley 1994:343]"
NB:  zero-sum game
(116)  "The result is of prime significance.  For whenever the actual trade volumes are less than the general equilibrium ones, agent society is not extracting all the welfare from trade that it might.  If the agents could coordinate their activities beyond their local nieghborhoods they could all be made better off.  here we see that _even though T produces exchanges that are nearly Pareto-optimal locally, the resulting market has far from optimal welfare properties globally."
(120)  "In what follows we make our agents more human, first, by giving them finate lines and, second, by permitting then preferences to evolve.  We shall see that the effect of these ew rules is to add variance to the distribution of prices and to modigy the price itself.  In fact, the mean price will follow a kind of 'random walk.'"
(123)  "Overall, the effet of tadie is to _further skew_ the distribution of wealth in society.  So, while trade increases the carrying capacity allowing more agents to survive, it also increase the inequality of the wealth distribution.  In this sense, there is a tradeoff between economic equality and economic performance."
(127)  "To economists, envrionmental pollution is the classic negative externality.  Externalities are important since their existence is an indication that an economuy is not achieving efficient resource allocation."
(129)  "In this case [of pollution transients] the artificial economy requires roughly twice as long to recover its statistical price equilibrium as it did to deviate from it."  
(136)  "Certain economists ascribe nearly magical powers to markets.  Markets are idealized to operate frictionlessly, without central authority, costlessly allocating resources to their most efficient use.  In this world of complete decentralization and Pareto efficiency, the only possible effect of government intervention is to 'gum up' the perfect machinery.  While this extreme view is perhaps little more than a caricature - and few would admit to holding it in toto - it is also, unfortunately, a postion frequently promulgated in policy circles, especially when there is no econometric or other evidence upon which to base decisionmaking."
(137)  "From the computational evidence above, we think that there is good reason to be skeptical of the predominant forcus on fixed-point equilibria.  Economies of autonomous adaptive agents - and of humans - may be far from equilibrium systems.  And, in turn, far from equilibrium economics might well turn out to be far richer than equilibrium economics."
(142)  "Recall from the discussion of cultural tags in Chapter III that the number of bitwise disagreements between two strings is the Hamming distance between them. So, the immune system - call it I _ searches itself for the substring closest to D in Hamming distance."
(153)  "The main point of the preceding chapters is simply this:  _A wide range of important social, or collective, phenomena can be made to emerge from the spatio-temporal interaction of autonomous agents operating on landscapes under simple local rules_."
"When the textbook assumptions were relaxed, the economy was pushed further from equilibrium.  From a policy standpoint, the analysis raises deep questions as to the allocative efficiency of unregulated markets."
(154)  "Over the preceeding chapters, we have built agent rules of movement, sexual reproduction, cultural transmission, group membership, trade, inheritance, credit, immune response, and disease propagation."
(160)  "But the mere fact that trade _can have_ the effects displayed above suggest that economic policy _can be_ a kind of population policy.  At the very least, artificial societies raise important policy questions.  And they may help answer some fo them."
(162)  "Mathematical social science generally adopts two assumptions.  First, agents can have different utility functions, but each agent's is _fixed_.  Second, every one is assumed to be doing the _same_ thing with its particular utility function - for example, maximizing it or maximizing its expected value."
(163)  "The deep point is that _our rules create the cliffs we drive off_."
footnote 6  "A particularly powerful example of this occurs in the model of Ackley and Littman[1992] in which they populate thier artificial world with a group of very capable but fiercely competitive agents and find that the population stays small and eventually goes extinct.  when less capable agents are released into the same environment, their population rises and lasts indefinitely."
(164)  "In all of the foregoing discussion we have been talking about autonomous adaptive agents interacting in, and with, a completely artificial environment, that is, one that follows rules that we devise.  It might be instructive to put the agents in a 'real' environment - that is to say, a physically realistic environmental model.  When the agents emit a pollutant it might be fed into an air-quality model, or groundwater toxification code, which would feed back into the agents' subsequent behavior, and so on.  Such systems could be quite useful in alerting us to counterintuitive, nonlinear effets (good and bad) of various regulatory policies or technological changes."

"With artificial agents in a simple environmental model it might be possible to grow a history that mimics the true history of some ancient tribe as it migrated in response to environmental changes.  At the time of this writing, we are collaborating with archeologists at the Santa Fe Institute and the Tree Ring Laboratory at the University of Arizona on a project whose aim is to grow the population dynamics and settlement patterns of the Anasazi from 400 to 1400 AD in the Long House Valley area of Black Mesa, using environmental and demographic data reconstructed by archeological methods."
(170)  "As a final example, then, we distribute agent preferences for like neighbors uniformly between 25 percent and 50 percent.  This adds more tolerant individuals to the previous run.  Will the final picture more closely resemble the modestly segregated outcome of animation VI-5, or the more completely segregated society of animation VI-6?  Animation VI-7 provides the answer.

"Adding this degree of tolerance is _not sufficient to generate_ desegration - indeed a highly segregated pattern endures.

"Now, had you been _shown_ this terminal pattern - the already emerged phenomenon - of segregation, you might well have concluded that virtually every agent had demanded that all neighbors be of its color.  Not so!  The question, then, is this:  How _little_ racism is enough to 'tip' a society into this segregated pattern?  In turn, is racial segregation reversible through 'invasion' by a handful of 'color-blind' individuals?  How much does it take to 'tip' things the other way?  Do these simple models help explain why we are so often surprised by the _true_ preference landscapes - for example, the Serb-Bosnian one - that burst forth when suppressive institutions are suddenly dismantled?

"If we want to understand political change, we need ways to study the 'match' between political institutions and the underlying preference landscapes.  When the match is good, there is political stability, but when the match is bad, frustration accumulates and there can be sudden releases of conflictual energy."
(177)  "Clearly, agent-based social science does not seem to be either deductive or inductive in the usual senses.  But then what is it?  We think _generative_ is an appropriate term.  The aim is to provide initial microspecifications (initial agents, environments, and rules) that are _sufficient to generate_ the macrostructures of interest.  We consider a given macrostructure to be 'explained' by a given microspecification when the latter's generative sufficiency has been established.  As suggested in Chapter I, we interpret the question, 'can you explain it?' as asking 'can you grow it?'  In effect, _we are proposing a generative program for the social sciences and see the artificial society as its principal scientific instrument_."
(176)  "In some areas, it may be that simulation _really is_ the best we can do."
(181)  "...20,000 lines of code that make up Sugarscape."

No comments:

Post a Comment