Using subsimulations¶
Summary¶
In this example, we use sub-simulations and plot different parts of the data set against each other.
Describing Subsimulations¶
To use subsimulations means that you run different simulations which differ in specific sets of parameters. There are often situations where this is a good approach. Sometimes just because it structures your process, sometimes for technical reasons. We use a good example for a technical reason here: Two simulations that use different implementations of the Prisoners Dilemma. Such a dilemma is described by four parameter settings and here, we do not want all of the possible combinations between the two dilemmas (4*4=16). We want only two combinations - the two dilemmas.
To describe subsimulations, you can have some more .conf files next to stosim.conf, one for each simulation.
In those, you can specify parameter settings that only hold in those simulations. Subsimulations inherit
settings from the main configuration stosim.conf and their settings overwrite
settings there. The subsim example mentions its subsimulations in
stosim.conf in a section called [simulations], like so:
# -------------------------------------
# An example StoSim configuration file
# For a simpler example with more comments, see the basic example
# -------------------------------------
[meta]
name: Sub-simulation Example
maintainer:Nicolas Hoening
[control]
executable:./main.py
runs:20
scheduler:fjd
fjd-interval:.2
[params]
# Parameters to your simulation
# if you want more than one setting for a parameter, give a comma-separated list
n:100
epochs:200
# how cooperative are the non-learners?
mean_coop:0.2,0.8
# what ratio of all agents is learning?
ratio_learning:0.25,0.75,1
# we define two sub-simulations here
[simulations]
configs:sim1,sim2
Note
We als have some more parameter settings than in the basic example.
n is the number of agents, ratio_learning indicates how many agents
will change their likelihood to cooperate based on their experiences and
mean_coop is the mean likelihood to cooperate of the
(non-learning) agents.
Have a look in the executable for this simulation to see what the simulation is doing
exactly. Basically, we are interested in the likelihood to
cooperate which the learning agents will arrive at (they start
at around 0.5).
The second dilemma (in sim2.conf) is technically not a Prisoners Dilemma.
I just played with the numbers to see how the very very simple
learning algorithm I used behaves when in the dilemma temptation
and reward (as well as penalty and suckers’ payoff) are the same.
In this case, we need to provide the configuration files
sim1.conf and
sim2.conf. Note that we left out the
.conf-extension when we mentioned them.
In the subsimulations, we define a unique subset of settings in their own [params] - section.
This is where we describe the different outcomes of interactions (payoff-wise)
in our two versions of the Prisoner’s Dilemma.
We can also give each simulation an own name and name a new maintainer for each subsimulation.
Here is the configuration (figures left out for now) in sim1.conf:
[meta]
name: Usual Prisoners Dilemma
[params]
pd_t:5
pd_r:3
pd_p:1
pd_s:0
And here is sim2.conf:
[meta]
name: Skewed Prisoners Dilemma
maintainer: Kramer
[params]
pd_t:5
pd_r:5
pd_p:1
pd_s:1
To make it clear: It does not appear any different to your executable if a parameter setting is defined in one configuration file
or the other. If you specified a comma-separated list of parameter values in stosim.conf or in a sub-simulation config file,
your executable code will get one config file with all parameters available in
the params section (the configuration file for each setting is also put into the subfolder in the data dir).
You only need to know that settings in the sub-simulation config files (here: sim1.conf and sim2.conf) overwrite
the general settings in the main configuration file (stosim.conf).
This is one of the ways in which I believe StoSim makes your life easier :)
Running¶
Let’s run this thing. This is the output I get:
nic@fidel:/media/data/projects/stosim/trunk/examples/subsim$ stosim --run
********************************************************************************
Running simulation Subsimulation Example
********************************************************************************
[fjd-recruiter] Hired 1 workers in project "Sub-simulation_Example".
[fjd-dispatcher] Started on project "Sub-simulation_Example"
[fjd-dispatcher] Found 1 job(s) and 1 free worker(s).....
[fjd-dispatcher] No (more) jobs.
[fjd-recruiter] Fired 1 workers in project "Sub-simulation_Example".
The second fjd-dispatcher line changes as it goes on. In the beginning it had 240 jobs. If you use the pbs scheduler, your output looks different.
Afterwards, you’ll find that the subfolders of the examples/subsim/data directory will now also carry the simulation name.
This is the listing of that directory after the simulation is run:
nic@fidel:/media/data/projects/stosim/trunk/examples/subsim$ ls data
simsim1_mean_coop0.2_n100_epochs200_pd_t5_ratio_learning0.25_pd_p1_pd_s0_pd_r3 simsim2_mean_coop0.2_n100_epochs200_pd_t10_ratio_learning0.25_pd_p1_pd_s0_pd_r5
simsim1_mean_coop0.2_n100_epochs200_pd_t5_ratio_learning0.75_pd_p1_pd_s0_pd_r3 simsim2_mean_coop0.2_n100_epochs200_pd_t10_ratio_learning0.75_pd_p1_pd_s0_pd_r5
simsim1_mean_coop0.2_n100_epochs200_pd_t5_ratio_learning1_pd_p1_pd_s0_pd_r3 simsim2_mean_coop0.2_n100_epochs200_pd_t10_ratio_learning1_pd_p1_pd_s0_pd_r5
simsim1_mean_coop0.8_n100_epochs200_pd_t5_ratio_learning0.25_pd_p1_pd_s0_pd_r3 simsim2_mean_coop0.8_n100_epochs200_pd_t10_ratio_learning0.25_pd_p1_pd_s0_pd_r5
simsim1_mean_coop0.8_n100_epochs200_pd_t5_ratio_learning0.75_pd_p1_pd_s0_pd_r3 simsim2_mean_coop0.8_n100_epochs200_pd_t10_ratio_learning0.75_pd_p1_pd_s0_pd_r5
simsim1_mean_coop0.8_n100_epochs200_pd_t5_ratio_learning1_pd_p1_pd_s0_pd_r3 simsim2_mean_coop0.8_n100_epochs200_pd_t10_ratio_learning1_pd_p1_pd_s0_pd_r5
Plotting with reduced data-sets¶
With all the data we have, we also should make some figures to look at - we’ll use our parameter settings to plot different data sets, such that we can meaningfully compare the outcomes of different settings.
Note
We tell StoSim about these figures in the subsimulation configuration file, as opposed to the main file stosim.conf (we could also have
done that). If figures are described in subsimulations configs, StoSim will select only the data that were generated within the settings of this
subsimulation when it creates these figures. I find that pretty convenient :)
Figure 1 shows likelihood to cooperate of learners when the non-learners form a
cooperative environment (mean_coop:0.8) and Figure 2 does this in a
non-cooperative environment (mean_coop:0.2). Figure 3 shows payoffs of
everyone in both environments.
Figure descriptions can also be put in the subsimulation’s config files. This should help to keep bigger projects a bit structured (just start numbering at 1 in each file).
This is from sim1.conf:
[figure1]
name: simulation1_cooperative
y-range: [0:1]
x-label: iteration
y-label: cooperation_probability
plot1: _name:learners-minority, _ycol:4, _type:line, ratio_learning:0.25, mean_coop:0.8
plot2: _name:learners-majority, _ycol:4, _type:line, ratio_learning:0.75, mean_coop:0.8
plot3: _name:learners_all, _ycol:4, _type:line, ratio_learning:1, mean_coop:0.8
[figure2]
name: simulation1_non-cooperative
y-range: [0:1]
x-label: iteration
y-label: cooperation_probability
plot1: _name:learners-minority, _ycol:4, _type:line, ratio_learning:0.25, mean_coop:0.2
plot2: _name:learners-majority, _ycol:4, _type:line, ratio_learning:0.75, mean_coop:0.2
plot3: _name:learners-all, _ycol:4, _type:line, ratio_learning:1, mean_coop:0.2
[figure3]
name: simulation1_payoff
y-range: [0:7]
x-label: iteration
y-label: payoff
plot1: _name:non-learners_in_coop, _ycol:2, _type:line, sim:sim1,ratio_learning:0.25, mean_coop:0.8
plot2: _name:learners_in_coop, _ycol:3, _type:line, ratio_learning:0.75,mean_coop:0.8
plot3: _name:non-learners_in_non-coop, _ycol:2, _type:line, ratio_learning:0.25, mean_coop:0.2
plot4: _name:learners_in_non-coop, _ycol:3, _type:line, ratio_learning:0.75, mean_coop:0.2
This is from sim2.conf:
[figure1]
name: simulation2_cooperative
y-range: [0:1]
x-label: iteration
y-label: cooperation_probability
plot1: _name:learners-minority, _ycol:4, _type:line, ratio_learning:0.25, mean_coop:0.8
plot2: _name:learners-majority, _ycol:4, _type:line, ratio_learning:0.75, mean_coop:0.8
plot3: _name:learners_all, _ycol:4, _type:line, ratio_learning:1, mean_coop:0.8
[figure2]
name: simulation2_non-cooperative
y-range: [0:1]
x-label: iteration
y-label: cooperation_probability
plot1: _name:learners-minority, _ycol:4, _type:line, ratio_learning:0.25, mean_coop:0.2
plot2: _name:learners-majority, _ycol:4, _type:line, ratio_learning:0.75, mean_coop:0.2
plot3: _name:learners-all, _ycol:4, _type:line, ratio_learning:1, mean_coop:0.2
[figure3]
name: simulation2_payoff
y-range: [0:7]
x-label: iteration
y-label: payoff
plot1: _name:non-learners_in_coop, _ycol:2, _type:line, ratio_learning:0.25, mean_coop:0.8
plot2: _name:learners_in_coop, _ycol:3, _type:line, ratio_learning:0.75, mean_coop:0.8
plot3: _name:non-learners_in_non-coop, _ycol:2, _type:line, ratio_learning:0.25, mean_coop:0.2
plot4: _name:learners_in_non-coop, _ycol:3, _type:line, ratio_learning:0.75, mean_coop:0.2
Here are the plots we get - Since we use only line plots, all data was averaged. First the probability to cooperate:
Payoffs for agents in sub-simulation example
And the payoffs the agents got:
Payoffs for agents in sub-simulation example
Note
Again, the example is here to show you functionality rather than to convey scientific value :) However, we can see in simulation 1 that - using our simple learning behaviour - cooperation decreases heavily among learners, no matter if they started in a cooperative or non-cooperative environment, or if learners were in the majority or not - and they are not able to extract higher profits overall. This is different when the dilemma is not really a Prisoners Dilemma, but reward and temptation are the same. Cooperation started around 0.5 and basically stays the same. However, there is an interesting bump at the beginning of the simulation which we can call an orientation phase …