Ultimatesim

Agent based simulations of iterated ultimatum game.

The ultimatum game is a simple two player game. In the game the offerer makes an offer to the responder (in the current implementation the offer will be in 0-10 in integer values) and subsequently the responder either accepts the offer or rejects the offer [0,1]. If the offer is accepted both receive what was offered, if not, both players receive no payoff.

The tendency of humans players to a) offer fair shares, and b) decline unfair offers has been taken as a demonstration of "irrational behavior". Recent work shows however that --- in evolutionary simulations --- fairness evolves in single iterations of the game (see: http://www.pnas.org/content/110/7/2581.full).

The current software aims at making it easy to experiment with different offering and response strategies in iterated games (hence, players in a population can) meet the same players again. The current software allows one to specify different policies for offering and responding; possibly policies that change as a function of their historical succes. After creating a population in which certain response and offer policies are present with pre-specified rates one can run multiple simulations of iterated games and inspect their behavior.

The software

The software was primarily written to support our own investigations, but due to its structure can easily be expanded and used. Below we detail the basic usgage of the software, provide some minimal examples, and describe how one can build on the current software.

Running a simulaiton

After downloading all the files in this repository, running a simulation is easy and, if one has python 3, numpy, and matplotlib installed, can be done by typing:

python main.py

in your terminal after navigating to the ultimatesim folder.

Now, this is not too interesting, but it give you a gist of a small simulation and produces a few plots.

For those just wish to use the framework without modifications the above command together with changes made to the config.py file should satisfy. The examples below detail the usage of the config.py file.

Some examples

To specify the parameters of a simulation please make changes to the config.py file. Most of the content of that file should be self explanatory, but for completeness we provide an exmple here. We setup a simulation by first setting up the population of agents:

#### First, we define the population
# Population size (e.g., the number of agents in the population:
popSize = 100

# An array describing the offering policies in the population:
# see the /policy folder for the options
offerPolicies = np.array(["REMOffer", "RANDOMOffer"], dtype=str)

# An array describing the response policies in the population:
responsePolicies = np.array(["REMResponse", "RANDOMResponse"], dtype=str) 

# A matrix (who's entries should sum to 1) describing the probabiltiy
# of occurancy of different strategy combination in the population
# Columns denot offer policies, rows denote response policies
probMatrix = np.matrix([[.7,0.1],[0.1,0.1]])

Note that this generates a population with 100 agents of which (in expectation) 70% of the agents have the "Rational Economic Man" (REM) offering policy (e.g., always offering 1 assuming that it will be accepted), and the REM response policy. About 10% of the agents have a REM offering policy, but a (uniformly) random response strategy (and so on). This provides a fairly flexible way of setting up populations with agents following different strategies.

After setting up the popultation we setup the details of the simulation; this should be relatively self-explanatory:

### Next, we specify the simulations
# Simulation length; the number of rounds played
# per simulation
simLength = 100

# Number of sims; the total number of simulations to run with this setting:
nSims = 100

And finally, we determine the output of this simulations:

# Create outut plots (put this to true, otherwise nothing really happens ;))
createPlots = True    
    
# Store all rund to /data folder as csv?
# This could generate a lot of data if you run many long simulations
storeAll = False

Running

python main.py

now produces (amongst others) the following plots:

• 
• 
• 

And from there on one can start experimenting with different policies, in different proportions, etc. etc.

Creating and agent

Now the above use is still not very intersting, the more interesting part of the library is the ability to create new agents that determine their behavior based on historical response. This can be done be creating a new offer or repsonse policy in the /policies/ folder, and inheriting from the base offer and response policies. For examples, please see the currently implemented policies.

We will add a list describing the available agents shortly.

Further expanding the software

We are quite well aware that we are missing a bunch of nice to have and neat to have functionality, but we are trying to push this forward gradually. Things we would like to add still:

• More overview statistics
• Flexible rules of the game
• More policies
• Evolution / fittness and birth / death of agents
• etc. On the last point it is interesting to see that evoution (e.g, birht-death of agents based on fittness) seems very similar to "learning" agents in repeated games.

More info & questions

For more info and contact details see: http://www.nth-iteration.com