intro tutorial: Analysing model reporters #1955
Conversation
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## main #1955 +/- ##
=======================================
Coverage 79.45% 79.45%
=======================================
Files 15 15
Lines 1285 1285
Branches 285 285
=======================================
Hits 1021 1021
Misses 225 225
Partials 39 39 ☔ View full report in Codecov by Sentry. |
|
I set the PR to ready for review so that I can read the rendered tutorial on the RTD CI. |
|
@rht There's a |
|
Maybe the runtime is just too long for Read the Docs |
|
While you are at it. Can you also include a Also I find it suboptimal that another warning is raised, a FutureWarning regarding AgentSet. I think its a bit overwhelming for first-timers, especially since the model doesn't make any direct use the AgentSet feature (only indirectly in the schedulers) |
|
Also also: Can you fix the reference to the matplotlib installation in the beginning? The tutorial uses seaborn. And we have to install seaborn manually for the colab version, since its not a requirement of mesa |
| "outputs": [], | ||
| "source": [ | ||
| "# Create a point plot with error bars\n", | ||
| "g = sns.pointplot(data=results_filtered, x=\"N\", y=\"Gini\", linestyle='none')\n", |
There was a problem hiding this comment.
This seems to be slightly duplicated from the previous paragraph
g = sns.scatterplot(data=results_filtered, x="N", y="Gini")
g.set(
xlabel="Number of agents",
ylabel="Gini coefficient",
title="Gini coefficient vs. number of agents",
);Describing both seems to be an exercise in showcasing the feature of Seaborn, because they are essentially showing the same info. But if I have to choose which one to keep, the sns.pointplot with error bars seems to be more informative.
| { | ||
| "cell_type": "markdown", | ||
| "source": [ | ||
| "In this case it looks like the Gini coefficient increases slower for smaller populations. This can be because of different things, either because the Gini coefficient is a measure of inequality and the smaller the population, the more likely it is that the agents are all in the same wealth class, or because there are less interactions between agents in smaller populations, which means that the wealth of an agent is less likely to change." |
There was a problem hiding this comment.
The Gini coefficient being the way it is, is due to preferential attachment. The inequality scales faster than linearly against the number of nodes. I don't think this paragraph has incorporated preferential attachment in the explanation.
There was a problem hiding this comment.
Bit in doubt. Sometimes it's useful to show the same thing in two different ways to engrain the idea (in this case that it's just a dataframe that can be plotted in different ways).
There was a problem hiding this comment.
Did you mean to reply to #1955 (comment) ? If you say so, you have to explicitly say it's a different way to visualize the same data.
I assume it's because AgentSet is slightly slower than the previous version without it? |
With less agents the model runs are quicker and so the batch_run method is finished earlier. The resulting data is also more interesting.
Add a steps_not_given agent reporter to the model that's used in the batch_run() function. This way we agent plots can be discussed.
EwoutH
force-pushed
the
tutorial_improvement_2024
branch
from
428c25a to
07ce5b9
Compare
|
Okay fixed all the stuff (see commit messages), ready for another round of review! |
I quickly ran 2.2 against 2.1.5 for a few example models yesterday. We indeed sacrificed a bit of performance for the convenience of AgentSet. |
|
I updated the PR description. It renders correctly in Readthedocs. Please review and merge. Feel free to squash while merging! |
|
Note to self: Need to talk about nested / multidimensional aggregation somewhere. Like how to aggregate over multiple agents, over multiple iterations (and possibly over multiple timesteps). When to do what and how to do it properly. (after having seen students just throwing all there data on a big pile and drawing a single average out of it. And then reporting a confidence interval which means god knows what) Ideally maybe a full stack “how to perform experiments and report results with agent-based models”. Including:
And the theory and how to do it in Mesa sandwiched. |
Some enhancements to the Mesa Introductory Tutorial:
steps_not_given, which tracks the number of consecutive steps an agent hasn't transacted. This addition enriches the tutorial's analytical depth, demonstrating how to handle multiple reporters.It also fixes Readthedocs not having enough runtime to finish running the notebook, by focussing the batch_run on runs with fewer agents (which run faster).
Tested and renders correctly in Readthedocs. Please review and merge. Feel free to squash while merging.