Duffy's ABM and human experiments

Experimental

Behavioural

Economics

image courtesy US NIST

• limited numbers of subjects at once
• difficulty approximating real world incentives
• difficulty in controlling for characteristics of subjects
• difficult to avoid, e.g. "strategic playing" rather than desired behaviour

but

• empirically based
• guaranteed to approximate real human learning behaviour and rationality.. since they are human

"Zero Intelligence Agents"

due to Gode and Sunder (1993)

Unconstrained Robot Traders

Human and robot tests

Robot Traders with a bank balance

Human traders

courtesy Bibliodessy http://bibliodyssey.blogspot.com/2007/12/turk-chess-automaton-hoax.html

It turns out you can do a reasonable job of approximating human behaviour with agents acting randomly constrained by institutions...

Duffy, John, 2006 "Agent-based models and human subject experiments"

"Mechanical Turk", and stock broker Turing tests, and plain old-fashioned statisical comparison of outcomes

"Learning" agents

Reinforcement learning

image courtesy http://chomskyscolorlessgreenideas.blogspot.com/2009/09/chinese-room.html

...of individual and aggregate behaviour

Belief-based

http://bigeyedeer.wordpress.com/2008/05/12/this-cartoon-is-a-dogs-breakfast/

Classifier systems

Pantazi et al. BMC Medical Informatics and Decision Making 2004 4:19 doi:10.1186/1472-6947-4-19

Replicator Dynamics

Genetic algorithms

The Duplicators by Murray Leinster (1964)

http://xkcd.com/720/

Replicator Dynamics

Genetic programming

Missing: Neural networks

Evolutionary algorithms

image courtesy http://www.genetic-programming.org/

Agent-based

Computational

Economic

Models

image courtesy Renee Ting http://www.reneesbookoftheday.com/2006/10/invention-of-hugo-cabret-by-brian.html

Replicator Dynamics

• huge number of agents
• unbounded time scale
• freedom from boredom for agents
• "growing" a system is a form of insight

but

• learning rules may be too simple to easily calibrate against "real" behaviour
• no necessary reason that behaviours are any less arbitrary or over simple than e.g. general equilibrium economic models - still "armchair theorising"