Learning to Be Thoughtless in Python with Mesa

I recently read Who's Counting? Uniting Numbers and Narratives with Stories from Pop Culture, Puzzles, Politics, and More by John Allen Paulos. In the Chapter 6 section titled "Voting Blocs: Red States, Blue State, and a Model for Thoughtless Voting," Paulos mentions an interesting paper by Joshua Epstein, Learning To Be Thoughtless: Social Norms And Individual Computation. In this paper, Epstein proposes a simple but compelling agent-based model for how social norms evolve in a community. This model is interesting because it captures two fundametal forces that motivate human behavior quite elegantly:

1. the social influence the behavior of others exerts on our own, and
2. laziness.

Paulos uses Epstein's model as a potential explanation for the increasing polarization of American Politics, which, while plausible, is not the focus of this post. After following the citation to Epstein's paper, I thought it would be fun and instructive to implement this model in Python and reproduce the main results of the paper.

Epstien's paper uses the decision to drive on the left- or right-hand side of the road as the social norm under analysis. For the purpose of this post, we will use a boolean flag norm to indicate whether or not an agent conforms to the social norm (defined, for instance, by driving on the right-hand side of the road in the United States) or does not (driving on the left-hand side of the road in the United States).

Epstein names his agents "lazy statisticans." These lazy statisticians determine whether or not they follow the social norm based on the norm-following behavior of the other agents in their vicinity. The size of the neighborhood that influences a statistician's behavior is their radius. With this terminology, we can begin to implement lazy statisticians in Python. Throughout this post we use Mesa, an agent-based modeling framework that

allows users to quickly create agent-based models using built-in core components (such as spatial grids and agent schedulers) or customized implementations; visualize them using a browser-based interface; and analyze their results using Python’s data analysis tools.

Each lazy statistician is a Mesa Agent. Following the paper, if no initial value for norm is past, we choose one with a coin flip, and if no initial radius is passed, we choose one uniformly between one and fitfy.

When a lazy statistician updates their norm-following behavior, they do so based on the proportion of norm-following agents in their vicinity as determined by their radius (optionally offset by Δ).

When a lazy statistician updates their norm-following behavior, they do so in two steps. First they update the radius of neighboring agents that influence them, then they update their norm.

When updating their radius, a lazy statistician first compares the proportion of norm-followers in their current neighborhood, defined by their radius, to the proportion of norm-followers in the neighborhood defined by radius + 1. If these not close (for the purposes of this post, we mostly follow the guideline of the paper and define "close" to be within 5% of each other), the lazy statistician increases their radius by one, since more data is better. If they are close, the lazy statistician compares the proportion of norm-followers in their current neighborhood to the proportion of norm-followers in the neighborhood by radius - 1. If these are close, the lazy statisician decreases their radius by one (due to their laziness).

After updating their radius, a lazy statistician updates their norm to agree with the majority of the agents in their neighborhood (including themselves). Note that including themsevles means that there are an odd number of agents (2 * radius + 1) agents in their neighborhood, so their will never be an ambiguous case where exactly 50% of agents in the neighborhood are norm-followers.

If there are $N$ agents, at each time step $N$ agents are drawn randomly with replacement and their status is updated as described above. Mesa does not include this sort of bootstrap scheduler out-of-the-box, but it is simple to implement.

With LazyStatistician and BoostrapActivation defined, we are ready to implement our model.

The most important feature of of this model is that the agents are placed on a circle (as indicated by the torus=True keyword argument to SingleGrid.) Another point of note is that the constructor take the agent class as a parameter. This flexibility will be useful in reconstructing examples from the paper that add noise and/or temporal shocks to the lazy statistician's update rules.

With ThoughtlessModel in hand, we can start reproducing Epstein's computational results. The paper presents six runs of the model, exploring

stylized facts regarding the evolution of norms: [l]ocal conformity, global diversity, and punctuated equilibria.

Run 1. Monolithic Social Norm, Individual Computing Dies Out¶

Following Epstein, for the first simulation we set each lazy statistician's initial norm to False. In this case, we anticipate that norm should remain False for all agents for all time and each agent's radius should shrink to the minimum value of one.

We choose a seed value for reproducibility.

Instead of simulating 190 lazy statisticians as in the paper, we simulate 500. We find that this yields interesting outcomes in more complex situations far more often (as the seed varies). As in the paper, we simulate 275 time steps.

Visualizing the evolution of each agent's norm and radius over time confirms our expectations.

Though the paper does show this scenario, we get an analagous result when all of the initial norms are True.

In each of these cases, "individual computing dies out" is reflected by the fact that every agent's radius eventually decreases to one, the smallest possible.

Run 2. Random Initial Norms, Individual Computing At Norm Boundaries¶

Run one verified our intuition in the basic but unrealistic case of a uniform ("monolithic") initial value of norm. In run two we explore the more interesting and realistic situation where the initial values of norm are random.

We see that there are two disjoint regions where norm is True and two where it is False. (Recall that the agents are on a circle so the regions at the left and right ends of this plot are the same from their perspective.) "Individual computing at norm boundaries" corresponds to the fact that all agents sufficiently far in the interior of the constant-norm regions eventually have radius one, and it is only agents near the borders of these regions that "compute" in the sense of consulting more than their direct neighbors in their norm update decisions.

In order to understand the factors influencing the width of these edges that compute, we reduce the tolerance used when updating an agent's radius from 5% to 2.5%.

We see that a lower tolerance causes an agent to expand their radius more often (remember the agent checks a larger radius before considering a smaller radius) and therefore we end up with more agents with nontrivial radius. In the opposite direction, doubling the tolerance to 10% causes agents to increase their radius less often and therefore the boundary where computing occurs shrinks.

Run 4. Modest Noise Level and Endogenous Neighborhood Norms¶

For implementation reasons we will replicate run three in the paper last and skip to run four now. The first two scenarios treated a noiseless case where agents never spontaneously change their norm regardless of the input from their neighbors. This noiselessness is unrealistic, so we introduce a NoisyLazyStatistician subclass of LazyStatistician that randomly draws a new norm occasionally according to its noise parameter.

We simulate behavior when agents randomize their norm due to noise 15% of the time and visualize the results below.

These results are visually quite interesting but not hugely different from those in previous runs. For the most part we get two blocks each for each value of norm. Interestingly, in this case small endogenous dissenting islands occasionally apear inside these blocks and manage to persist for a few timesteps before disappearing.

Run 5. Higher Noise and Endogenous Neighborhood Norms¶

The next simulation doubles the noise level from 15% to 30%.

At this higher noise level, we see the bocks of consistent norm erode over time, and the endognenous islands that appear persist for much longer than in the previous simulation.

Run 6. Maximum Noise Does Not Induce Maximum Search¶

In the previous two noisy scenarios, the average agent's radius was consistently higher than in the noiseless scenario, as shown below.

This observation prompts the questions: how much does the average radius increase as the noise increases? Does it increase without bound in the 100% noise case? We answer these questions in the next simulation.

We see that the average radius does not increase without bound even with maximum noise, which is interesting.

Run 3. Complacency in New Norm¶

Finally we return to run three from the paper, which we have deferred until after the noisy scenarios for implementation reasons. In this run, most of the time our agents behave noiselessly, but for ten time steps for $130 \leq t < 140$ the system experiences a "shock" which causes each agent to behave completely randomly. We first implement this behavior in the ShockedLazyStatistician subclass of NoisyLazyStatistician.

We now simulate this scenario and visualize the results.

Quite interestingly, many new blocks of consistent norm appear almost immediately after the shock subsides. This behavior is likely due to the fact that an agent's radius cannot increase all that much during the ten time step shock (as was shown in run five.) This explanation is consistent the the following plot of average agent radius over time.

Though not present in the paper, we consider two additional scenarios. The first is adding low-level noise (15%) to all time steps in the above scenario.

The results are not surprising given the outcomes of models 3a and 4.

For our last scenario we maintain this noise level, but increase the number of steps by a factor of ten and shock the system whenever $130 \leq t \operatorname{mod} 275 < 140.$

The results are quite fascinating but not unexpected.

We have now reproduce all of the scenarios from Epstein's Learning To Be Thoughtless: Social Norms And Individual Computation. The behavior of these models is quite interesting and I certainly learned quite a bit about agent-based modeling using Mesa.

This post is available as a Jupyter notebook here.