Download presentation
Presentation is loading. Please wait.
Published byShanon Stewart Modified over 8 years ago
1
Experiments with Minimally Intelligent Agents and Minimal Institutions: Structure and Behavior Shyam Sunder, Yale University Barcelona LeeX Experimental Economics Summer School in Macroeconomics Universitat Pompeu Fabra Barcelona, June 9, 2016
2
10/2/2016Sunder, Structure and Behavior2 Humanities and Science Science does not know its debt to imagination. Ralph Waldo Emerson Vivisection is a social evil because if it advances human knowledge, it does so at the expense of human character. George Bernard Shaw The theoretical broadening which comes from having many humanities subjects on the campus is offset by the general dopiness of the people who study these things. Richard P. Feynman (Nobel Laureate in Physics) Economics has an amazing capacity to summarize staggeringly complex phenomena by the application of only a handful of principles Charles R. Plott
3
10/2/2016Sunder, Structure and Behavior3 Overview Origin of experimental economics in examination of aggregate phenomena Gradual, steady shift towards micro-levels due to –Analytical process and reasoning –Incremental research questions –Unlike assumption in theory, people do not optimize well by intuition Today, much experimental work has shifted to examination of individual behavior and of economies populated by artificial agents Shift to individual behavior has accentuated the ever-present dilemma of social sciences in trying to be a science on one hand, and handle humans at the same time What are the antecedents and consequences of this trend? Usefulness of organizing experimental economics into three streams: –Structural: macro properties of social structures –Behavioral: behavior of individuals, and –Agent-based: exploration of links between the micro and macro phenomena At least the structural part of economics can be firmly rooted in the tradition of sciences, bypassing the free-will dilemma of social sciences
4
10/2/2016Sunder, Structure and Behavior4 Examining Market Institutions Chamberlin (1948) examined the behavior of a market institution under controlled conditions of his classroom Vernon Smith (1962), a subject of Chamberlin) redesigned and systematically varied the market conditions to examine price, allocation, and extraction of surplus Both designs deviated significantly from Walrasian tatonnement abstraction; they fleshed them out with details, using stock market as a guide –Economic environment (market demand and supply) and market design as independent variables –Market level outcomes as dependent variables
5
10/2/2016Sunder, Structure and Behavior5 Data from Experiments Experiments can yield a great deal of data Data are limited only by interest and imagination of the experimenter, and ingenuity in capturing data without distracting subjects from their task in a significant way Chamberlin gathered three pieces of data for each transaction (price, seller cost and buyer value), and the transaction sequence Examples of data he did not gather: the clock time of transactions, details of the bargaining process (time elapsed, price proposals, number of proposals, number of counter-parties bargained with), etc.
6
10/2/2016Sunder, Structure and Behavior6 Data to Meet Experimental Goals Most experiments can yield a great deal of data We gather only what we need in order to address the question(s) we wish to answer on the basis of the experiment Constraints: –Technology of data gathering, eased by development of computer technology to conduct economics experiments) –The possibility of interaction between data capture and subjects Given Chamberlin’s goals, asking subjects to report their transactions immediately after they completed each transaction served his purposes well, causing little interference with subjects’ trading
7
10/2/2016Sunder, Structure and Behavior7 Shift Towards Micro Phenomena Focus of experimental economics has gradually shifted from aggregate market level phenomena towards individual behavior Three factors seem to drive this shift –The logic of analytical method –Incremental research designs –Empirical finding that people, acting by intuition alone, are not good at optimization as typically assumed in derivation of equilibria in economic theory
8
10/2/2016Sunder, Structure and Behavior8 Logic of Analytical Method It is rare for the correspondence between the predictions of the relevant theory, and experimental data, to be nil or total If the experimenter has no or low expectation of correspondence between the two, observation of even a moderate relationship is seen as half full glass of water However, most experiments are designed to examine specific theories that have some legitimate prior claim to predictive power In such situations, any imperfections of correspondence between data and theory are seen as half empty, not half full, glass of water Seeking a fuller explanation to close the gap between data and theory is a natural reaction of most investigators
9
10/2/2016Sunder, Structure and Behavior9 Search for Higher Explanatory Power Following this logic, analysis and discussion of most experiments ends in a search for ways to increase the correspondence between data and theory Better prediction and explanation is the currency of scientific progress We look for ways to modify the model to enhance its explanatory power through analysis—breaking the problem down into progressively smaller components This logical pursuit shifts research question(s) to the next level of detail causing “micro-nization” of economics Discarding the details, to step back and see the big picture, is a less common reaction
10
10/2/2016Sunder, Structure and Behavior10 Demand, Supply and Experiments Simple economic theory: point of intersection of demand and supply determines price and allocations Economists’ deep faith in theory Neither Chamberlin’s nor Smith’s data corresponded precisely to the theory Smith saw half full glass of water, while Chamberlin saw the half empty part and set out to build a model to better explain the residual variation left unexplained by the simple demand- supply model (instantaneous demand/supply)
11
10/2/2016Sunder, Structure and Behavior11 Chamberlin (1948), Figure 3
12
10/2/2016Sunder, Structure and Behavior12 Smith (1962) Chart 1
13
10/2/2016Sunder, Structure and Behavior13 Incremental Research Designs A good part of our research (including experimental) is incremental, originating in proposals to gather data about some additional aspect of behavior, or additional analysis of existing data We make conjectures about how such data or analysis might help explain residual variation Incremental work dominates graduate seminars focused on critique and replication of extant work Easy to think of additional observations, motivations, and information conditions associated with individual participants to improve the fit between data and model
14
10/2/2016Sunder, Structure and Behavior14 Change in Models and Questions Both analytical logic and incremental pursuits change the model used Additional variables use up some degrees of freedom, but observations at micro-level are far more numerous than at macro-level Shift to micro level also changes the research question(s) being asked –“Why is the price equal x?” might be replaced by “why did trader y bid z?”
15
10/2/2016Sunder, Structure and Behavior15 Individual Behavior and the Dilemma of Social Sciences This shift towards micro-behavior confronts economics with a fundamental dilemma shared among the social sciences As a science, we seek general laws that apply everywhere at all time, emulating physics, chemistry and biology Perfecting the scope and power of general laws of human behavior also implies squeezing out the essence of humanity—our free will What does it mean to have a science of individual human behavior?
16
10/2/2016Sunder, Structure and Behavior16 Free Will Free will, independent thinking, and ability to choose are essential to our concept of self We believe in our power and ability to do what we wish, beyond what is predictable on the basis of our circumstances, beliefs, and tendencies Ability to rise above our circumstances as the essence of human identity We can choose deliberately, in ways unpredictable to others Else, we would slip to the status we assign to animals, plants and inanimate objects
17
10/2/2016Sunder, Structure and Behavior17 Humanities: Eternal Truths Humanities celebrate infinite variety of human behavior, but no laws of behavior In epics and literature: eternal verities, but no laws of behavior –Epics (Mahabharata, Iliad) Duryodhana, Yudhishtira, Arjuna –Literature (Dante’s Inferno, Shakespeare’s Hamlet) Human truths, questions, and tendencies repeated through history, always with a new twist People choose in ways unpredictable on the basis of their circumstances Celebration of infinite variation in human nature
18
10/2/2016Sunder, Structure and Behavior18 Science: Eternal Laws Identifying laws of nature valid everywhere and all the time Essence: regularities of nature captured in known and knowable relationships among observable elements (including stochastic) Helps understand, explain, and predict If I know X, can I form a better idea of whether Y was, is or will be? Objects of science have no free will –A photon does not pause to enjoy the scenery –A marble rolling down the side of a bowl does not wonder about how hot the oil at the bottom is
19
10/2/2016Sunder, Structure and Behavior19 Social Science: Irresistible Force Meets Immovable Object Free will essential to our concept of self Without the freedom to act, we would be no different than a piece of rock Yet, the object of study in social science is us As a science, it must look for eternal laws that apply to humanity But stripped of freedom to act, and subject to such laws, there can be no humanity
20
10/2/2016Sunder, Structure and Behavior20 Mismatch of Science and Personal Responsibility Objects of science can have no personal responsibility They do not choose to do anything They are merely driven by their circumstances, like a piece of paper blown by gusts of wind, or a piece of rock rolling down the hill under force of gravity in the path of an oncoming car Or, perhaps an abused child who grows up to be an abusive parent, sans personal responsibility Science and personal responsibility do not mix well
21
10/2/2016Sunder, Structure and Behavior21 Neither Fish Nor Fowl This problem of social science is exemplified in the continuing attempts to build a theory of choice From science end: axiomatization of human choice as a function of innate preferences. People choose what they prefer How do we know what they prefer? Look at what they choose The circularity between preferences and choice might be avoided if there were permanency and consistency in preference-choice relationship across diverse contexts One could observe choice in one context, tentatively infer the preferences from these observations, and assuming consistent preferences, predict choice in other contexts Unfortunately, half-a-century of research has yielded little predictability of choice from inferred preferences across contexts (Friedman and Sunder 2004; Friedman et al. 2014) Individual human behavior appears to be unmanageably rowdy for scientists to capture in a stable set of laws While humanists may not take delight at such disappointments, but they can hardly be surprised (if they pay any attention to choice theory)
22
10/2/2016Sunder, Structure and Behavior22 Dilemma of Social Science Do we abandon free will, personal responsibility, and special human identity; and treat humans like other objects of science? That is, drop the “social” and become a plain vanilla science Or, do we abandon the search for universal laws, embrace human free will and unending variation of behavior, and join the humanities Either way, there will be no social science left Is there a way to keep “social” and “science” together in social science?
23
10/2/2016Sunder, Structure and Behavior23 Isolating Three Streams of Work Perhaps there is no general solution to this dilemma The dilemma does, however, point to the potential value of isolating streams of work where it may be more or less of a problem Significant parts of social sciences, and a large part of economics, are concerned with aggregate level outcomes of socio-economic institutions Institutions themselves do not need to be ascribed intentionality or free will Characteristics of the institutions can be analyzed by methods of science without running into these dilemmas This will leave analysis of individual behavior in the territory between science and humanities Agent-based models (in economics and elsewhere) could serve the bridging function between aggregate and individual phenomena Let us consider these possibilities
24
10/2/2016Sunder, Structure and Behavior24 Individuals I do not have much to add on the most complex problem of examining individual behavior It seems that we shall continue to examine ourselves and our behavior using both humanities as well as science perspectives, without ever reconciling the two into a single logical structure There seems to be no way out
25
10/2/2016Sunder, Structure and Behavior25 Institutions Experimental economics started out as investigation of aggregate level outcomes of market institutions using human subjects Attention has gradually shifted from aggregate outcomes to micro behavior –Logic of analytical approach –Incremental research designs A third reason is that predictions of aggregate outcomes (equilibrium analysis) are typically made assuming optimization by individuals Cognitive psychology showed that individuals are not very good at optimization by intuition This mismatch between the optimization assumption actual behavior at individual level has given additional impetus to “micro-nization” of experimental economics Thanks to recent findings using agent-based methods, we can conduct the study of social-economic institutions using methods of science
26
Background Gary S. Becker (JPE 1962) Distinguish rationality at market and individual levels Negatively-sloped demand curves result from the change in opportunities sets alone, largely independent of the decision rule; derivable from maximizing as well as random behavior 10/2/2016Decoupling Markets26
27
Background Vernon L. Smith (JPE 1962) In a double auction Profit motivated human traders Only 10-20 traders Observed market outcomes close to competitive equilibrium (derived from maximization, Walrasian auction, and atomistic competition) 10/2/2016Decoupling Markets27
28
Background Gode and Sunder (JPE 1993): serendpitously combined –Double auction market institution (without randomization in Smith), and –Budget constrained random choice (without market institution in Becker) 10/2/2016Decoupling Markets28
29
Market Behavior 1 10/2/2016Decoupling Markets29
30
Market Behavior 2 10/2/2016Decoupling Markets30
31
Market Behavior 3 10/2/2016Decoupling Markets31
32
10/2/2016Sunder, Structure and Behavior32 What Makes the Difference
33
Background Gode and Sunder (JPE 1993): Combine –Yields market outcomes approximate predictions of competitive equilibrium –Allocative efficiency of markets populated by “zero-intelligence” traders is high, and comparable to efficiency of markets populated with profit-motivated human traders 10/2/2016Decoupling Markets33
34
Background Plott and Sunder (JPE 1982) examined asset markets with uncertain, heterogeneous payoffs, and asymmetric information In double auction markets information can be disseminated from informed to uninformed traders to yield approximate rational expectations equilibria as measured by prices and allocations 10/2/2016Decoupling Markets34
35
Decoupling Markets35 Equilibria in an Asset Market State XState Y Prob.=0.4Prob.=0.6 Trader TypeExpected Dividend I0.40.10.22 II0.30.150.21 III0.1250.1750.155 PI Eq. Price Asset Holder 0.4 Trader Type I Informed 0.22 Trader Type I Uninformed RE Eq. Price Asset Holder 0.4 Trader Type I 0.175 Trader Type III 10/2/2016
36
Decoupling Markets36
37
Present Research Question Do the results from markets without uncertainty (Becker, Smith, Gode and Sunder) extend to markets with uncertainty and asymmetric information populated by minimally-intelligent traders? –Note that profit-motivated human traders converged near RE equilibria in Plott and Sunder (1982) markets –But we do not know how and why –Attaining RE equilibria calls for making the right conjectures –We cannot directly observe human trading strategies –So we replace human traders by trading algorithms to unpack the aspects of behavior that may be sufficient to yield theoretically derived equilibria 10/2/2016Decoupling Markets37
38
Design of Algorithmic Traders Populate double auctions with simple algorithmic traders with two characteristics –Means-end heuristic to adjust their aspiration levels –Zero-intelligence bids and offers relative to the current aspiration levels 10/2/2016Decoupling Markets38
39
Means-End Heuristic Use Newell and Simon (The Simulation of Human Thought, 1959) –Given a current state and a goal state, an action is chosen which will reduce the difference between the two. The action is performed on the current state to produce a new state, and the process is recursively applied to this new state and the goal state Initial aspiration –For the uninformed: expected payoff –For the informed: pay off under the known state Adjustment of aspiration levels with each observed price P(t) using parameter α ~U(0.05, 0.5) –ASP (t+1) = (1-α) ASP (t) + α P(t) 10/2/2016Decoupling Markets39
40
Zero-Intelligence Bids and Offers Bids: ~ U(0, ASP) Offers:~U(ASP, 1) Double Auction Rule –At the start of each period –Set current bid = 0, and current ask = 1 –A higher bid replaces current bid, and a lower ask replaces current ask –As soon as a current bid equals or exceeds current ask, a transaction is recorded at the bid/ask submitted first 10/2/2016Decoupling Markets40
41
Algorithmic Flow At time = 0 Set CAL = Expected Value of Dividend for the trader CAL i = Pr(X) * (DX i ) + Pr(Y) * (DY i ) Transaction Price Observed (p t ) After each transaction CAL Updated Based on Prior CAL and Observed Transaction Price CAL t+1 = (1- ) (CAL t ) + ( P t ) Inside Trade r No Set CAL = its dividend for the realized state Yes
42
No Current Bid = @ Bid ≥ Ask Is # ≤ 0.5 Randomly Choose Number # ~U[0,1] Is @ > Current Bid? Yes No Trade Occurs Yes Is @ < Current Ask Current Ask = @ Yes No Randomly Choose Number @ ~U[0,CAL] Randomly Choose Number @ ~U[CAL, 1] Algorithmic Flow
43
Results Plott and Sunder (1982) results of markets with profit-motivated human traders shown in blue Simulated results of same markets populated by minimally-intelligent algorithmic traders shown in red (cloud of 50 independent simulations and the median, α ~U(0.05, 0.5) drawn for each simulation) RE equilibrium price in green line Walrasian equilibrium price in broken line Also shown, comparisons of mean squared deviation, trading volume, and allocative efficiency 10/2/2016Decoupling Markets43
44
Decoupling Markets44 Equilibria in an Asset Market State XState Y Prob.=0.4Prob.=0.6 Trader TypeExpected Dividend I0.40.10.22 II0.30.150.21 III0.1250.1750.155 PI Eq. Price Asset Holder 0.4 Trader Type I Informed 0.22 Trader Type I Uninformed RE Eq. Price Asset Holder 0.4 Trader Type I 0.175 Trader Type III 10/2/2016
45
Decoupling Markets45
46
10/2/2016Decoupling Markets46
47
10/2/2016Decoupling Markets47
48
10/2/2016Decoupling Markets48
49
10/2/2016Decoupling Markets49
50
10/2/2016Decoupling Markets50
51
10/2/2016Decoupling Markets51
52
10/2/2016Decoupling Markets52
53
10/2/2016Decoupling Markets53
54
10/2/2016Decoupling Markets54
55
10/2/2016Decoupling Markets55
56
10/2/2016Decoupling Markets56
57
10/2/2016Decoupling Markets57
58
10/2/2016Decoupling Markets58
59
10/2/2016Decoupling Markets59
60
10/2/2016Decoupling Markets60
61
10/2/2016Decoupling Markets61
62
10/2/2016Decoupling Markets62
63
10/2/2016Decoupling Markets63
64
10/2/2016Decoupling Markets64
65
10/2/2016Decoupling Markets65
66
10/2/2016Decoupling Markets66
67
10/2/2016Decoupling Markets67
68
10/2/2016Decoupling Markets68
69
10/2/2016Decoupling Markets69
70
10/2/2016Decoupling Markets70
71
10/2/2016Decoupling Markets71
72
10/2/2016Decoupling Markets72
73
10/2/2016Decoupling Markets73
74
10/2/2016Decoupling Markets74
75
10/2/2016Decoupling Markets75
76
10/2/2016Decoupling Markets76
77
10/2/2016Decoupling Markets77
78
10/2/2016Decoupling Markets78
79
10/2/2016Decoupling Markets79
80
10/2/2016Decoupling Markets80
81
10/2/2016Decoupling Markets81
82
10/2/2016Decoupling Markets82
83
10/2/2016Decoupling Markets83
84
10/2/2016Decoupling Markets84
85
10/2/2016Decoupling Markets85
86
10/2/2016Decoupling Markets86
87
10/2/2016Decoupling Markets87
88
10/2/2016Decoupling Markets88
89
10/2/2016Decoupling Markets89
90
10/2/2016Decoupling Markets90
91
10/2/2016Decoupling Markets91
92
10/2/2016Decoupling Markets92
93
10/2/2016Decoupling Markets93
94
10/2/2016Decoupling Markets94
95
10/2/2016Decoupling Markets95
96
10/2/2016Decoupling Markets96
97
10/2/2016Sunder, Structure and Behavior97 Optimization and Equilibrium The standard approach of economic analysis has been to assume that individuals choose actions by optimizing given their preferences, information and opportunity sets Interaction of individual actions in the context of institutional rules yield outcomes (e.g., prices and allocations), equilibrium outcomes being of special interest Equilibrium predictions derived from assuming individual rationality could be suspect when such rationality assumption is not valid Agent-based simulations suggest that individual rationality can be sufficient but not necessary for attaining equilibria in the context of specific market institutions
98
10/2/2016Sunder, Structure and Behavior98 Why Equilibrium without Individual Optimization Why do the markets populated with simple budget-constrained random bid/ask strategies converge close to Walrasian prediction in price and allocative efficiency No memory, learning, adaptation, maximization, even bounded rationality Search for programming and system errors did not yield fruit Modeling and analysis supported simulation results
99
10/2/2016Sunder, Structure and Behavior99 Inference Perhaps it is the structure, not behavior, that accounts for the first order magnitude of outcomes in competitive settings Computers and experiments with simple agents opened a new window into a previously inaccessible aspect of economics Ironically, it was not through computers’ celebrated optimization capability Instead, through deconstruction of human behavior –Isolating the market level consequences of simple or arbitrarily chosen classes of individual behavior modeled as software agents
100
ZI Applications Non-binding price controls (Gode and Sunder 2004) Learning competitive equilibrium (Crockett, Spear and Sunder 2008) Three minimal market institutions (Huber, Shubik and Sunder 2010) An economy with personal currency (Angerer et al. 2010) Huange Farmer 10/2/2016Sunder, Structure and Behavior100
101
10/2/2016Sunder, Structure and Behavior101 Optimization Principle In physics: marbles and photons “behave” but are not attributed any intention or purpose Yet, optimization principle has proved to be an excellent guide to how physical and biological systems as a whole behave –At multiple hierarchical levels--brain, ganglion, and individual cell— physical placement of neural components appears consistent with a single, simple goal: minimize cost of connections among the components. The most dramatic instance of this "save wire" organizing principle is reported for adjacencies among ganglia in the nematode nervous system; among about 40,000,000 alternative layout orderings, the actual ganglion placement in fact requires the least total connection length. In addition, evidence supports a component placement optimization hypothesis for positioning of individual neurons in the nematode, and also for positioning of mammalian cortical areas. –(Makes you wonder what went wrong with human design when you see all the biases and incompetence of human cognition. –Could it be just the wrong benchmark?) Questions about “forests” and questions about “trees”
102
10/2/2016Sunder, Structure and Behavior102 Optimization Principle Imported into Economics Humans and human systems as objects of economic analysis Conflict between mechanical application of optimization principle and our self-esteem (free will) Optimization principle interpreted as a behavioral principle, shifting focus from aggregate to individual behavior Cognitive science: we are not good at optimizing Willingness among economists to abandon the optimization principle
103
10/2/2016Sunder, Structure and Behavior103 Dropping the “Infinite Faculties” Assumption Conlisk: –Empirical evidence in favor of bounded rationality –Empirical evidence on importance of bounded rationality –Proven track record of bounded rationality models (in explaining individual behavior) –Unconvincing logic of unbounded rationality All these reasons focus on the “trees” not “forest”
104
Simon on Reductionism The approach used here suggests that some important properties of market institutions are not simple translation of individual level behavior Spirit of Sonnenschein-Mantel-Debreu Theorem: the excess demand function for an economy is not restricted by the usual rationality restrictions on individual demands in the economy Herbert Simon, after originating the rationality critique in 1950s, and documenting bounded rationality of humans, also realized the necessity of decoupling aggregate and individual behavior “This skyhook-skyscraper construction of science from the roof down to the yet unconstructed foundations was possible because the behavior of the system at each level depended on only a very approximate, simplified, abstracted characterization of the system at the level next beneath. This is lucky, else the safety of bridges and airplanes might depend on the correctness of the ‘Eightfold Way’ of looking at elementary particles.” ( The Sciences of the Artificial,3 rd edition, p. 16) 10/2/2016Decoupling Markets104
105
Equilibria as Basins of Attraction Indeed, the powerful results of economic theory have been derived from “a very approximate, simplified, abstracted characterization of the system at the level next beneath.” Simon and Selten’s bounded rationality of individual behavior, and Smith and Plott’s markets may not be mutually inconsistent, even under uncertainty, information asymmetry, and heterogeneous preferences The “economic man” is maligned by a misunderstanding of its scientific purpose and role Economics borrowed optimization from physics as a structural principle to identify basins of attraction (not as a behavioral description) Adam Smith may not have gone far enough; a large part of economic efficiency is generated by common-sense constraints on economic choices made in the context of institutions evolved for social interaction 10/2/2016Decoupling Markets105
106
Beautiful Experiments in Science After Crick and Watson reported the discovery the double helix structure, theorists and empiricists raced to the search for the code that RNA uses to translate information in DNA to synthesize 20 amino acids and their peptide chains or proteins It was a hopelessly complex problem, until Nirenberg tried a breathtakingly simple test tube manipulation: Let us see what amino acid an artificial RNA sequence that repeats the same letter again and again (UUUUU… which does not appear in nature) produces? It produces phenylalanine. The code began to crumble and was soon deciphered: each of the 64 possible codons or “words” consisted of three “letters” each produced one of the 20 amino acids. Examples of other masters of beautiful experiment in biology: Severo Ochoa, Matthew Meselson, Franklin Stahl, and Joshua Lederberg 10/2/2016Decoupling Markets106
107
Thank You Shyam.sunder@yale.edu www.som.yale.edu/faculty/sunder
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.