Software Agents in Economic Environments Robert S. Gazzale Ph.D. Candidate, Department of Economics Jeffrey MacKie Mason Professor, Dept. of Economics & School of Information CARAT 2002/2003 April 9, 2003
Dec 21, 2015
Software Agents in Economic Environments
Robert S. GazzalePh.D. Candidate, Department of Economics
Jeffrey MacKie MasonProfessor, Dept. of Economics & School of Information
CARAT 2002/2003
April 9, 2003
April 9, 2003 CARAT: Gazzale & MacKie Mason
Collaborators Chris Brooks Computer Science, University of San Francisco
Yan Chen School of Information
Rajarshi Das IBM Institute for Advanced Commerce
Ed Durfee AI Lab, EECS
Jeff Kephart IBM Institute for Advanced Commerce
April 9, 2003 CARAT: Gazzale & MacKie Mason
A Model of Economic Modeling: Alternative View
Environment
Outcome
April 9, 2003 CARAT: Gazzale & MacKie Mason
A Model of Economic Modeling:Which Mapping?
Environment
Outcomes
April 9, 2003 CARAT: Gazzale & MacKie Mason
Equilibrium: The Mapping from Environment to Outcome
One Agent Environment Optimal action
Non-cooperative Games Nash Equilibrium
Given what everybody else is doing, no one agent can change strategy to do better.
April 9, 2003 CARAT: Gazzale & MacKie Mason
Problem: Finding Equilibria
Is it solvable? If so, will agents find it? Bounded Rationality (Herbert Simon)
Cognition is not free. Satisfice rather than optimize?
April 9, 2003 CARAT: Gazzale & MacKie Mason
Problem: Out of equilibrium matters!
Particularly if agents are boundedly rational Do we get to equilibrium? What happens on path to equilibrium?
April 9, 2003 CARAT: Gazzale & MacKie Mason
More Problems with “Equilibrium”
Which Equilibrium? If there are many equilibria, which is
going to happen when?
April 9, 2003 CARAT: Gazzale & MacKie Mason
Why Software Agents?
Useful in alleviating equilibrium issues. Cheap. Present/Future of software agents in real markets Particularly where equilibrium not
solvable.
April 9, 2003 CARAT: Gazzale & MacKie Mason
Application 1: Convergence to Equilibrium
Nash
Equilibrium
Theory: Supermodular (SPM) games played by learning agents converge to Nash Equilibrium
April 9, 2003 CARAT: Gazzale & MacKie Mason
Application 1: Convergence to Equilibrium
Nash
Equilibrium
Theory: Supermodular (SPM) games played by learning agents converge to Nash Equilibrium
Trust me, you don’t need to know what this is!
April 9, 2003 CARAT: Gazzale & MacKie Mason
Application 1: Convergence to Equilibrium
Nash
Equilibrium
Theory makes no predictions if NOTSupermodular.
April 9, 2003 CARAT: Gazzale & MacKie Mason
Application 1: Convergence to Equilibrium
Nash
Equilibrium
Is more supermodular “better”?
April 9, 2003 CARAT: Gazzale & MacKie Mason
Application 1: Convergence to Equilibrium
Nash
Equilibrium
Answers to these questions important in designing markets!
April 9, 2003 CARAT: Gazzale & MacKie Mason
Convergence to Equilibrium: Human Experiment
Methodology Design game where parameter
controls whether or not game is SPM Laboratory experiments with human
subjects playing for real money!
April 9, 2003 CARAT: Gazzale & MacKie Mason
Convergence to Equilibrium: Human Experiment
Methodology Design game where parameter
controls whether or not game is SPM Laboratory experiments with human
subjects playing for real money!
April 9, 2003 CARAT: Gazzale & MacKie Mason
Convergence to Equilibrium: Human Experiment Results
Problem: Dynamics not complete with human subject experiments (60 rounds)
April 9, 2003 CARAT: Gazzale & MacKie Mason
Convergence to Equilibrium: Human Experiment Results
Efficiency of Outcome:Experimental Data
40%
50%
60%
70%
80%
90%
100%
0 10 20 30 40 50 60 Round
Fracti
on
of
Maxim
um
Welf
are
a20b00a20b18a20b20a20b40a10b20
April 9, 2003 CARAT: Gazzale & MacKie Mason
Convergence to Equilibrium: Human Experiment Results
Efficiency of Outcome:Experimental Data
40%
50%
60%
70%
80%
90%
100%
0 10 20 30 40 50 60 Round
Fracti
on
of
Maxim
um
Welf
are
a20b00a20b18a20b20a20b40a10b20
Will this treatment catch-up?
April 9, 2003 CARAT: Gazzale & MacKie Mason
Convergence to Equilibrium: Human Experiment Results
Efficiency of Outcome:Experimental Data
40%
50%
60%
70%
80%
90%
100%
0 10 20 30 40 50 60 Round
Fracti
on
of
Maxim
um
Welf
are
a20b00a20b18a20b20a20b40a10b20
Will this treatment catch-up?
Will any of these pull ahead?
April 9, 2003 CARAT: Gazzale & MacKie Mason
Software agents to complete dynamics
Methodology Select various learning models Endow agents with these models Calibrate models with actual data Compare calibrated learning models Endow pool of agents with best model
and let run!
April 9, 2003 CARAT: Gazzale & MacKie Mason
Software agents to complete dynamics
Use of computation power For each learning model and treatment
For each set of parameters (1100 sets) 12 agents play in each iteration for 60 rounds 1500 iterations of game
8,910,000,000 “decisions” in <6 hours! Select Parameters that most-closely fit
data. For best learning model
12 agents each iteration for 1000 rounds 1500 iterations of game
April 9, 2003 CARAT: Gazzale & MacKie Mason
Simulation ResultsSimulated Results
70%
75%
80%
85%
90%
95%
100%
50 150 250 350 450 550 650 750 Round
Fra
cti
on
of
Maxim
um
Welf
are
a20b00a20b18a20b20a20b40a10b20
“Never” does catch up!
Pulls ahead for a short while!
April 9, 2003 CARAT: Gazzale & MacKie Mason
Application 2: Agents Solving Difficult Problems
Many problems without analytical solution Natural domain for use of computer science methods to find optimum Many are “hill-climbing” methods
Economics needs to inform these solutions
April 9, 2003 CARAT: Gazzale & MacKie Mason
A not so hard problem for an agent . . .
No matter where we start, rather easy to get to the summit!
April 9, 2003 CARAT: Gazzale & MacKie Mason
A more difficult landscape . . .
Tough to get from here
to here
BundlePrice
Per-articlePrice
April 9, 2003 CARAT: Gazzale & MacKie Mason
A more difficult landscape . . . Made a little easier . . .
BundlePrice
Per-articlePrice
Use economic knowledge to:
reduce the search space!
April 9, 2003 CARAT: Gazzale & MacKie Mason
A more difficult landscape . . . Made a little easier . . .
select better starting values!Bundle
Price
Per-articlePrice
Use economic knowledge to:
April 9, 2003 CARAT: Gazzale & MacKie Mason
A more difficult landscape . . . Made a little easier . . .
FeePer-articlePrice
Use economic knowledge to:
April 9, 2003 CARAT: Gazzale & MacKie Mason
A more difficult landscape . . . Made a little easier . . .
FeePer-articlePrice
Use economic knowledge to:
supply gradient information!
April 9, 2003 CARAT: Gazzale & MacKie Mason
Problem 1: Pricing Problem
Firm Sells Information Goods Consumer demand uncertain Many different pricing schedules possible General rule: Higher profits from schedules that are harder to learn. What schedule? No analytical solution!
April 9, 2003 CARAT: Gazzale & MacKie Mason
The Pricing Problem: Results
adjusting Linear Two-part NonlinearBlock
Adaptive uses knowledge to move among schedules!
April 9, 2003 CARAT: Gazzale & MacKie Mason
Problem 2: Battle of the Agents
Highly complex environment Large Search space Actions of competitor warp my landscape. Result: Computer science algorithms, without economic knowledge, perform quite poorly
April 9, 2003 CARAT: Gazzale & MacKie Mason
Computer algorithm, no economic knowledge
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
x 104
0. 5
0
0.5
1
1.5
2
2.5
3
3.5
4x 10
4
Iteration
Average Profit per Iteration for Zero knowledge Producers
Equilibrium Profit
Equilibrium
April 9, 2003 CARAT: Gazzale & MacKie Mason
Computer algorithm, with economic knowledge
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
x 104
0
0.5
1
1.5
2
2.5
3
3.5
4x 10
4
Iterations
36000
zero-knowledge
gradient
tremble
noisenoise+adjacency
number of categories
equilibrium profit
Gradient info
}
Reduce search space
}
Better starting values