# This file introduces experiments with the knowledge renewal capability, modeled as a time window of memory.
# This experiment is reported in: Clark, J. (2016). When are Real-Time Feedback Loops Most Valuable? New Insights from Bandit Simulations of Decision Making in Turbulent Environments. Proceedings of HICSS-49.
# Its results are featured in figs. 7-9.
# For figs. 3-6, see clark_experiment_01.py
# Import statements: do not change
import random
import sys
from os import path
sys.path.insert(0, path.join(sys.path[0], "bandito"))
from bandito.banditexperiment import BanditExperiment
# Run using Python 3.4 on Windows 7, random seed 12345
random.seed(12345)
# Define the experiment:
# This one introduces memory; only needs two latency levels to calculate "cost of latency". 480 experiments.
BanditExperiment(strategy=[0.02, 0.25, 0.5, 0.75, 1],
turbulence=[0, 0.005, 0.01, 0.02, 0.04, 0.08, 0.16, 0.32],
latency=[0, 16],
memory=[50, 100, 150, 200, 250, 300, 350, 400, 450, 500],
replications=1000,
experiment_name="clark02").run()
from os import path sys.path.insert(0,path.join(sys.path[0],"bandito")) from bandito.banditexperiment import BanditExperiment # By setting a random number seed, you can try to guarantee that your results are repeatable. # Making note of your operating system and Python versions is also advisable. # Simply comment out the following two lines if you want new random results each time you run the program. # test experiments used random seed 12345 random.seed(12345) # Define the experiment: # Experimental treatments are defined by assigning lists of values to the arguments # of the BanditExperiment function. The simulation engine will iterate through the lists, # running the full number of simulation replications (500 by default) for each value. For # example, Posen & Levinthal's (2012) Figure #1 compared five levels of the strategy variable # tau from 0.02 to 1. To replicate their experiment, we assign strategy=[0.02,0.25,0.5,0.75,1] # and leave all other arguments with their defaults. # # Some other arguments do not set up experimental treatments but control the simulation parameters: # replications, arms, turns, debug, and experiment_name. The optional experiment_name determines the # name of the output files. If not assigned, the current datetime will be used. # # Default values are used for arguments not assigned values. # See bandito/banditexperiment.py to examine the default arguments. BanditExperiment(strategy=[0.02,0.25,0.5,0.75,1], replications=100, experiment_name="sample").run() # Run this file by typing something like: python sample_experiment.py # Output files will be found in the 'output' directory.
import sys from os import path sys.path.insert(0,path.join(sys.path[0],"bandito")) from bandito.banditexperiment import BanditExperiment # By setting a random number seed, you can try to guarantee that your results are repeatable. # Making note of your operating system and Python versions is also advisable. # Simply comment out the following two lines if you want new random results each time you run the program. random.seed(12345) # Define the experiment: # Experimental treatments are defined by assigning lists of values to the arguments # of the BanditExperiment function. The simulation engine will iterate through the lists, # running the full number of simulation replications (500 by default) for each value. For # example, Posen & Levinthal's (2012) Figure #1 compared five levels of the strategy variable # tau from 0.02 to 1. To replicate their experiment, we assign strategy=[0.02,0.25,0.5,0.75,1] # and leave all other arguments with their defaults. # # Some other arguments do not set up experimental treatments but control the simulation parameters: # replications, arms, turns, debug, and experiment_name. The optional experiment_name determines the # name of the output files. If not assigned, the current datetime will be used. # # Default values are used for arguments not assigned values. # See bandito/banditexperiment.py to examine the default arguments. BanditExperiment(strategy=[0.02,0.25,0.5,0.75,1], replications=25000, experiment_name="PLfig1").run() # Run this file by typing something like: python sample_experiment.py # Output files will be found in the 'output' directory.
from os import path
sys.path.insert(0, path.join(sys.path[0], "bandito"))
from bandito.banditexperiment import BanditExperiment
# By setting a random number seed, you can try to guarantee that your results are repeatable.
# Making note of your operating system and Python versions is also advisable.
# Simply comment out the following two lines if you want new random results each time you run the program.
random.seed(12345)
# Variations on the simulation to test the robustness of P+L 2012's figure 3 to different parameters
BanditExperiment(arms=[5],
turns=[100],
memory=[100],
strategy=[0.02, 0.25, 0.5, 0.75, 1],
turbulence=[0, 0.005, 0.01, 0.02, 0.04, 0.08, 0.16, 0.32],
replications=1000,
experiment_name="fig3_5arms100turns").run()
BanditExperiment(arms=[5],
turns=[2000],
memory=[2000],
strategy=[0.02, 0.25, 0.5, 0.75, 1],
turbulence=[0, 0.005, 0.01, 0.02, 0.04, 0.08, 0.16, 0.32],
replications=1000,
experiment_name="fig3_5arms2000turns").run()
BanditExperiment(arms=[20],
turns=[100],
memory=[100],
strategy=[0.02, 0.25, 0.5, 0.75, 1],
turbulence=[0, 0.005, 0.01, 0.02, 0.04, 0.08, 0.16, 0.32],
# This file replicates the experiment of Posen & Levinthal (2012), figure 3.
# Import statements: do not change
import random
import sys
from os import path
sys.path.insert(0, path.join(sys.path[0], "bandito"))
from bandito.banditexperiment import BanditExperiment
# By setting a random number seed, you can try to guarantee that your results are repeatable.
# Making note of your operating system and Python versions is also advisable.
# Simply comment out the following two lines if you want new random results each time you run the program.
random.seed(12345)
# Define the experiment:
BanditExperiment(strategy=[0.02, 0.25, 0.5, 0.75, 1],
turbulence=[0, 0.005, 0.01, 0.02, 0.04, 0.08, 0.16, 0.32],
replications=25000,
experiment_name="PLfig3").run()
# Warning: this took over 14 hours for my laptop to simulate. While testing, you may want to decrease the number of replications to fewer than 25000!
# This file replicates the basic bandit simulation (like P+L fig 1) at several
# levels of latency in the feedback loop, and several levels of turbulence. The
# goal is to show the "value of real-time feedback" or conversely "the cost of
# latency" at different levels of turbulence.
# This experiment is reported in: Clark, J. (2016). When are Real-Time Feedback Loops Most Valuable? New Insights from Bandit Simulations of Decision Making in Turbulent Environments. Proceedings of HICSS-49.
# Its results are featured in figs. 3-6.
# For figs. 7-9, see clark_experiment_02.py
# Import statements: do not change
import random
import sys
from os import path
sys.path.insert(0, path.join(sys.path[0], "bandito"))
from bandito.banditexperiment import BanditExperiment
# Run using Python 3.4 on Windows 7, random seed 12345
random.seed(12345)
# Define the experiment:
# 1000 reps seems a little more reasonable than P+L's 25000. This is 240 experimental treatments x 1000 replications and should run in about 6 hours on my laptop. Amended as clark01b with the initialization bias fix.
BanditExperiment(strategy=[0.02, 0.25, 0.5, 0.75, 1],
turbulence=[0, 0.005, 0.01, 0.02, 0.04, 0.08, 0.16, 0.32],
latency=[0, 1, 2, 4, 8, 16],
replications=1000,
experiment_name="clark01").run()