Man and Superman Human Limitations innovation and emergence

Man and Superman Human Limitations, innovation and emergence in resource competition Robert Savit University of Michigan

Collaborators Katia Koelle, Biology, University of Michigan Wendy Treynor, Psychology, UM Richard Gonzalez, Department of Psychology, UM Thanks to Yi Li, Physics, UM

Introduction • Theme of workshop is the design, prediction and control of collectives • Generally we understand the collectives to be composed of silicon agents, or at least related thereto. • But, many situations in which collectives may be involve carbon-based agents.

Introduction (2. ) • In particular, may be interested in collectives of humans, or collectives some of whose agents are human and some silicon. • Examples: – Markets and their regulation. – Systems in which humans exercise judgement or intervene in systems that are basically collectives of silicon agents. Eg. Logistics supply networks or networks of sensors and actuators which can be overridden by human controllers.

Introduction (3) • Many ways in which human agents different from silicon ones • lessons we learn from the study of collectives of silicon agents may have to be modified when we try to design, predict and control collectives of humans, or mixed collectives

Introduction (4) • Will report on preliminary controlled experiments with humans playing the minority game. • Indications of interesting new phenomena which may be important in design and control of human collectives

Outline • A. epistemological considerations: A story of psychologist-physicist collaboration • B. Simple review of the minority game with and without evolution. • C. Description of the experiments • D. Naive expectations of the outcome of human MG. • E. Expected limitations (and strengths) of human agents vis-à-vis silicon agents • F. What really happens • G. Why? ? ? • H. Directions for future work.

Outline • A. epistemological considerations: A story of psychologist-physicist collaborations • B. Simple review of the minority game with and without evolution. • C. Description of the experiments • D. Naive expectations of the outcome of human MG. • E. Expected limitations (and strengths) of human agents vis-à-vis silicon agents • F. What really happens • G. Why? ? ? • H. Directions for future work.

Epistemological Considerations • Never underestimate the naïveté of a psychologist nor the ignorance of a physicist • OR • Never underestimate the ignorance of a psychologist nor the naïveté of a physicist.

Outline • A. epistemological considerations: A story of psychologist-physicist collaborations • B. Simple review of the minority game with and without evolution. • C. Description of the experiments • D. Naive expectations of the outcome of human MG. • E. Expected limitations (and strengths) of human agents vis-à-vis silicon agents • F. What really happens • G. Why? ? ? • H. Directions for future work.

B. The Minority Game (a. k. a. El Farol Bar Problem) Resources Resource 0 Resource 1 Players • At each time step each player picks a resource each player on minority resource gets one point. Players in the majority group get nothing. • Objective: Each player wants to maximize his total points.

History of Minority Resources How do the players make their choices? t 1 2 3 4 5 6 7 8 9 Minority Resource 0 1 1 0 1 0 0 Window of last m minorities • EEach player has several (two) randomly generated strategies of window (memory) m. • AAt each step, the player uses the strategy that would have maximized its gains over the entire history. m=3 Strategy Next step, choose 0

An example of an m=3 strategy Recent History Predicted next minority group 000 0 001 010 011 100 101 110 111 1 1 0 0 0 1 1

The population of group 1 as a function of time (N=101) high standard deviation Poor use of the resource Low standard deviation Good use of the resource

Maladaptive behavior— poor systemwide performance Degrading performance---too much information Phase transition Emergent coordination of agent choice

The Minority Game with Evolution • Agents can change their strategies • Different rules – fixed m – variable m.

The Minority Game with Evolution--Results The σ2/N of two evolution games (all after 300 generations). The normal minority games have higher σ2/N. For comparison σ2/N for standard MG at dip . 07.

Outline • A. epistemological considerations: A story of psychologist-physicist collaborations • B. Simple review of the minority game with and without evolution. • C. Description of the experiments • D. Naive expectations of the outcome of human MG. • E. Expected limitations (and strengths) of human agents vis-à-vis silicon agents • F. What really happens • G. Why? ? ? • H. Directions for future work.

Description of the Experiments • N participants, each at a computer terminal • Each participant paid a flat sum plus $. 05 each time he is in the minority group (generally) • See the history of minority groups • See a running total of their winnings • No other information • 5 seconds to make each decision • Game runs for 400 time steps

Outline • A. epistemological considerations: A story of psychologist-physicist collaborations • B. Simple review of the minority game with and without evolution. • C. Description of the experiments • D. Naive expectations of the outcome of human MG. • E. Expected limitations (and strengths) of human agents vis-à-vis silicon agents • F. What really happens • G. Why? ? ? • H. Directions for future work.

What do Humans Do? 7± 2

What do Humans Do? • Therefore, expect best performance at Nc 19. Actually, finite size effects indicate a better value is Nc 15. • But maybe this is too naïve (or ignorant)…

Outline • A. epistemological considerations: A story of psychologist-physicist collaborations • B. Simple review of the minority game with and without evolution. • C. Description of the experiments • D. Naive expectations of the outcome of human MG. • E. Expected limitations (and strengths) of human agents vis-à-vis silicon agents • F. What really happens • G. Why? ? ? • H. Directions for future work.

Human Limitations and Strengths • Boredom • Memory limitations • Processing limitations – Biases – Systematic error in processing. Eg. overestimates of probabilities based on recent events – Random errors in processing – Emotions – Fallacies of causal inference—I. e. limitations in understanding about the way the system works

Human Limitations and Strengths (cont. ) • Possibility for great creativity – Possible source for response to non-stationarity or non-autonomy – Also possible weakness.

Outline • A. epistemological considerations: A story of psychologist-physicist collaborations • B. Simple review of the minority game with and without evolution. • C. Description of the experiments • D. Naive expectations of the outcome of human MG. • E. Expected limitations (and strengths) of human agents vis-à-vis silicon agents • F. What really happens • G. Why? ? ? • H. Directions for future work.

What Really Happens? • s 2/N as a function of time • s 2/N as a function of N • player performance as a function of strategy complexity

What Really Happens? • s 2/N as a function of time • s 2/N as a function of N • player performance as a function of strategy complexity

2/N as a function of time s

Comparison: Silicon vs. Carbon s 2/N as a function of time (N=5, 17, 101)

What Really Happens? • s 2/N as a function of time • s 2/N as a function of N • player performance as a function of strategy complexity

s 2/N as a function of N

s 2/N as a function of N • Note good performance (relative to RCG) for all N • Note oscillations • Need more data to determine s 2/N vs. N quantitatively (will come back to this)

What Really Happens? • s 2/N as a function of time • s 2/N as a function of N • player performance as a function of strategy complexity

Silicon Player Performance as a Function of Strategy Complexity • In evolutionary computer games, best performing agents have simplest strategies

Human Player Performance as a Function of Strategy Complexity

• Horizontal axis a measure of determinism of agent’s strategy, assuming m 3. • In fact, best performance is for m=0 strategies!! • Next best are m=1 strategies.

Back to s 2/N as a function of N • • 7 2 implies Nc 15 But, humans’ strategies seem to evolve. So, Maybe log 2 mt 7 2 In which case, s 2/N will be small for all N

Outline • A. epistemological considerations: A story of psychologist-physicist collaborations • B. Simple review of the minority game with and without evolution. • C. Description of the experiments • D. Naive expectations of the outcome of human MG. • E. Expected limitations (and strengths) of human agents vis-à-vis silicon agents • F. What really happens • G. Why? ? ? • H. Directions for future work.

Player Performance as a Function of Strategy Complexity • Why? ? – Appropriately clever (but not too clever) and insightful agents? – Boredom? • Is boredom an evolutionarily selected for adaptive strategy? ? – an adaptive mechanism which we evolved in order to limit our cleverness.

Outline • A. epistemological considerations: A story of psychologist-physicist collaborations • B. Simple review of the minority game with and without evolution. • C. Description of the experiments • D. Naive expectations of the outcome of human MG. • E. Expected limitations (and strengths) of human agents vis-à-vis silicon agents • F. What really happens • G. Why? ? ? • H. Directions for future work.

Future Work • Need continued close collaboration between social and natural scientists to bridge the gulf created by mutual ignorance and naïveté. • Need to develop thereby a richer epistemology of social dynamics than is now afforded either by social psychology/sociology or by econophysics.

Future Work (2) • Methods must include well designed and controlled experiments to better determine what the important underlying dynamics and principles are. • Examples – Need to understand what is the operative dynamics underlying simple strategy selection by humans – Top-down vs. emergent coordination—experiments in progress