## Finding Pi From Random Numbers

If I gave you 10,000 random numbers between 0 and 1, how precisely could you estimate pi? It’s Pi Approximation Day (22/7 in the European format), which seems like the perfect time to share some math!

When we’re modeling something complex and either can’t or don’t want to bother to find a closed-form analytic solution, we can find a way to get close to the answer with a Monte Carlo simulation. The more precisely we want to estimate the answer, the more simulations we would create.

One classic Monte Carlo approach to find pi is to treat our 10,000 numbers as the x and y coordinates of 5,000 points in a square between (0,0) and (1,1). If we draw a unit circle at (0,0), the percent of the points which land inside the circle gives us a rough estimate of the area – which should be pi / 4. Multiply by 4, and we have our estimate.

simulated mean:  3.1608
95% confidence interval: 3.115 3.206
confidence interval size: 0.091

This technique works, but it’s not the most precise. There’s a fair bit of variance – a 95% confidence interval is about .09 units wide, from 3.12 to 3.21. Can we do better?

Another way to find the quarter-circle’s area is to treat it as a function and take its average value. (Area is average height times width, the width is 1, so the area inside the quarter-circle is just its average height.) That would give us pi/4, so we multiply by 4 to get our estimate for pi.

$y^2+x^2=1$

$y^2=1-x^2$

$f(x)=\sqrt{1-x^{2}}$

We have 10,000 random numbers between 0 and 1; all we have to do is calculate f(x) for each and take the mean:

simulated mean:  3.1500
95% confidence interval: 3.133 - 3.167
confidence interval size: 0.0348

This gives us a more precise estimate;  the 95% confidence interval is less than half what it was!

But we can do better.

### Antithetic Variates

What happens if we take our 10,000 random numbers and flip them around? They’re just a set of points uniformly distributed between 0 and 1, so (1-x) is also a set of points uniformly distributed between 0 and 1. If the expected value of f(x) is pi with variance 0.1, then the expected value of f(1-x) should also be pi with variance of 0.1.

So how does this help us? It looks like we just have two different ways to get the same level of precision.

Well, if f(x) is particularly high, then f(1-x) is going to be particularly low. By pairing each random number with its converse , we can offset some of the error and get an estimation more closely centered around the true mean. Taking the average of two distributions, each with the same expected value should still give us the same answer.

(This trick, known as using antithetic variates, doesn’t work with every function, but works here because the function f(x) always decreases as x increases.)

simulated mean:  3.1389
95% confidence interval: 3.132 - 3.145
confidence interval size: 0.0131

Lo and behold, our 95% confidence interval has narrowed down to 0.013, still only using 10,000 random numbers!

To be fair, this only beats 22/7 about 30% of the time with 10,000 random simulations. Can we reliably beat the approximation without resorting to more simulations?

### Control Variates

It turns out we can squeeze a bit more information out of those randomly generated numbers. If we know the exact expected value for a part of the function, we can be more deliberate about offsetting the variance. In this case, let’s use c(x)=x^2 as our “control variate function”, since we know that the average value of x^2 from 0 to 1 is exactly 1/3.

Where our simulated function was

$f(x)=\sqrt{1-x^{2}}$

now we add a term that will have an expected value of 0, but will help reduce variance:

$f'(x)=\sqrt{1-x^{2}}+b(x^2-\frac{1}{3})$

For each of our 10,000 random x’s, if x^2 is above average, we know that f(x) will probably be a bit *below* average, and we nudge it up. If x^2 is below average, we know f(x) is likely a bit high, and nudge it down. The overall expected value doesn’t change, but we’re compressing things even further toward the mean.

The constant ‘b’ in our offset term determines how much we ‘nudge’ our function, and is estimated based on how our control variate covaries with the target function:

$\frac{Covariance(f(x), c(x))}{Variance(c(x))}$

(In this case, b is about 2.9) Here’s what we get:

simulated mean:  3.1412
95% confidence interval: 3.1381 - 3.1443
confidence interval size: 0.0062

See how the offset flattens our new function (in orange) to be tightly centered around 3.14?

This is pretty darn good. Without resorting to more simulations, we reduced our 95% confidence interval to 0.006.  This algorithm gives a closer approximation to pi than 22/7 about 57% of the time.

If we’re not bound by the number of random numbers we generate, we can get as close as we want. With 100,000 points, our control variates technique has a 95% confidence interval of 0.002, and beats 22/7 about 98% of the time.

These days, as computing power gets cheaper, we can generate 100,000 or even 1,000,000 random numbers with no problem. That’s what makes simulations so versatile – we can find ways to simulate even incredibly complicated processes and unbounded functions, deciding how precise we need to be.

Happy Pi Approximation Day!

(You may ask, why is there a “Pi Approximation Day” and not a “Pi Simulation Day”? Well, according to Nick Bostrom, every day is Simulation Day. Probably.)

## Which Cognitive Bias is Making NFL Coaches Predictable?

In football, it pays to be unpredictable (although the “wrong way touchdown” might be taking it a bit far.) If the other team picks up on an unintended pattern in your play calling, they can take advantage of it and adjust their strategy to counter yours. Coaches and their staff of coordinators are paid millions of dollars to call plays that maximize their team’s talent and exploit their opponent’s weaknesses.

That’s why it surprised Brian Burke, formerly of AdvancedNFLAnalytics.com (and now hired by ESPN) to see a peculiar trend: football teams seem to rush a remarkably high percent on 2nd and 10 compared to 2nd and 9 or 11.

What’s causing that?

His insight was that 2nd and 10 disproportionately followed an incomplete pass. This generated two hypotheses:

1. Coaches (like all humans) are bad at generating random sequences, and have a tendency to alternate too much when they’re trying to be genuinely random. Since 2nd and 10 is most likely the result of a 1st down pass, alternating would produce a high percent of 2nd down rushes.
2. Coaches are suffering from the ‘small sample fallacy’ and ‘recency bias’, overreacting to the result of the previous play. Since 2nd and 10 not only likely follows a pass, but a failed pass, coaches have an impulse to try the alternative without realizing they’re being predictable.

These explanations made sense to me, and I wrote about phenomenon a few years ago. But now that I’ve been learning data science, I can dive deeper into the analysis and add a hypothesis of my own.

The following work is based on the play-by-play data for every NFL game from 2002 through 2012, which Brian kindly posted. I spend some time processing it to create variables like Previous Season Rushing %, Yards per Pass, Yards Allowed per Pass by Defense, and QB Completion percent. The Python notebooks are available on my GitHub, although the data files were too large to host easily.

## Irrationality? Or Confounding Variables?

Since this is an observational study rather than a randomized control trial, there are bound to be confounding variables. In our case, we’re comparing coaches’ play calling on 2nd down after getting no yards on their team’s 1st down rush or pass. But those scenarios don’t come from the same distribution of game situations.

A number of variables could be in play, some exaggerating the trend and others minimizing it. For example, teams that passed for no gain on 1st down (resulting in 2nd and 10) have a disproportionate number of inaccurate quarterbacks (the left graph). These teams with inaccurate quarterbacks are more likely to call rushing plays on 2nd down (the right graph). Combine those factors, and we don’t know whether any difference in play calling is caused by the 1st down play type or the quality of quarterback.

The classic technique is to train a regression model to predict the next play call, and judge a variable’s impact by the coefficient the model gives that variable.  Unfortunately, models that give interpretable coefficients tend to treat each variables as either positively or negatively correlated with the target – so time remaining can’t be positively correlated with a coach calling running plays when the team is losing and negatively correlated when the team is winning. Since the relationships in the data are more complicated, we needed a model that can handle it.

I saw my chance to try a technique I learned at the Boston Data Festival last year: Inverse Probability of Treatment Weighting.

In essence, the goal is to create artificial balance between your ‘treatment’ and ‘control’ groups — in our case, 2nd and 10 situations following 1st down passes vs. following 1st down rushes. We want to take plays with under-represented characteristics and ‘inflate’ them by pretending they happened more often, and – ahem – ‘deflate’ the plays with over-represented features.

To get a single metric of how over- or under-represented a play is, we train a model (one that can handle non-linear relationship better) to take each 2nd down play’s confounding variables as input – score, field position, QB quality, etc – and tries to predict whether the 1st down play was a rush or pass. If, based on the confounding variables, the model predicts the play was 90% likely to be after a 1st down pass – and it was – we decide the play probably has over-represented features and we give it less weight in our analysis. However, if the play actually followed a 1st down rush, it must have under-represented features for the model to get it so wrong. Accordingly, we decide to give it more weight.

After assigning each play a new weight to compensate for its confounding features (using Kfolds to avoid training the model on the very plays it’s trying to score), the two groups *should* be balanced. It’s as though we were running a scientific study, noticed that our control group had half as many men as the treatment group, and went out to recruit more men. However, since that isn’t an option, we just decided to count the men twice.

## Testing our Balance

Before processing, teams that rushed on 1st down for no gain were disproportionately likely to be teams with the lead. After the re-weighting process, the distributions are far much more similar:

Much better! They’re not all this dramatic, but lead was the strongest confounding factor and the model paid extra attention to adjust for it.

It’s great that the distributions look more similar, but that’s qualitative. To do a quantitative diagnostic, we can take the standard difference in means, recommended as a best practice in a 2015 paper by Peter C. Austin and Elizabeth A. Stuart titled “Moving towards best practice when using inverse probability of treatment weighting (IPTW) using the propensity score to estimate causal treatment effects in observational studies“.

For each potential confounding variable, we take the difference in means between plays following 1st down passes and 1st down rushes and adjust for their combined variance. A high standard difference of means indicates that our two groups are dissimilar, and in need of balancing. The standardized differences had a max of around 47% and median of 7.5% before applying IPT-weighting, which reduced the differences to 9% and 3.1%, respectively.

So, now that we’ve done what we can to balance the groups, do coaches still call rushing plays on 2nd and 10 more often after 1st down passes than after rushes? In a word, yes.

In fact, the pattern is even stronger after controlling for game situation. It turns out that the biggest factor was the score (especially when time was running out.) A losing team needs to be passing the ball more often to try to come back, so their 2nd and 10 situations are more likely to follow passes on 1st down. If those teams are *still* calling rushing plays often, it’s even more evidence that something strange is going on.

Ok, so controlling for game situation doesn’t explain away the spike in rushing percent at 2nd and 10. Is it due to coaches’ impulse to alternate their play calling?

Maybe, but that can’t be the whole story. If it were, I would expect to see the trend consistent across different 2nd down scenarios. But when we look at all 2nd-down distances, not just 2nd and 10, we see something else:

If their teams don’t get very far on 1st down, coaches are inclined to change their play call on 2nd down. But as a team gains more yards on 1st down, coaches are less and less inclined to switch. If the team got six yards, coaches rush about 57% of the time on 2nd down regardless of whether they ran or passed last play. And it actually reverses if you go beyond that – if the team gained more than six yards on 1st down, coaches have a tendency to repeat whatever just succeeded.

It sure looks like coaches are reacting to the previous play in a predictable Win-Stay Lose-Shift pattern.

Following a hunch, I did one more comparison: passes completed for no gain vs. incomplete passes. If incomplete passes feel more like a failure, the recency bias would influence coaches to call more rushing plays after an incompletion than after a pass that was caught for no gain.

Before the re-weighting process, there’s almost no difference in play calling between the two groups – 43.3% vs. 43.6% (p=.88). However, after adjusting for the game situation – especially quarterback accuracy – the trend reemerges: in similar game scenarios, teams rush 44.4% of the time after an incomplete and only 41.5% after passes completed for no gain. It might sound small, but with 20,000 data points it’s a pretty big difference (p < 0.00005)

All signs point to the recency bias being the primary culprit.

## Reasons to Doubt:

1) There are a lot of variables I didn’t control for, including fatigue, player substitutions, temperature, and whether the game clock was stopped in between plays. Any or all of these could impact the play calling.

2) Brian Burke’s (and my) initial premise was that if teams are irrationally rushing more often after incomplete passes, defenses should be able to prepare for this and exploit the pattern. Conversely, going against the trend should be more likely to catch the defense off-guard.

I really expected to find plays gaining more yards if they bucked the trends, but it’s not as clear as I would like.  I got excited when I discovered that rushing plays on 2nd and 10 did worse if the previous play was a pass – when defenses should expect it more. However, when I looked at other distances, there just wasn’t a strong connection between predictability and yards gained.

One possibility is that I needed to control for more variables. But another possibility is that while defenses *should* be able to exploit a coach’s predictability, they can’t or don’t. To give Brian the last words:

But regardless of the reasons, coaches are predictable, at least to some degree. Fortunately for offensive coordinators, it seems that most defensive coordinators are not aware of this tendency. If they were, you’d think they would tip off their own offensive counterparts, and we’d see this effect disappear.

## Quantifying the Trump-iness of Political Sentences

You could say that Donald Trump has a… distinct way of speaking. He doesn’t talk the way other politicians do (even ignoring his accent), and the contrast between him and Clinton is pretty strong. But can we figure out what differentiates them? And then, can we find the most… Trump-ish sentence?

That was the challenge my friend Spencer posed to me as my first major foray into data science, the new career I’m starting. It was the perfect project: fun, complicated, and requiring me to learn new skills along the way.

To find out the answers, read on! The results shouldn’t be taken too seriously, but they’re amusing and give some insight into what might be important to each candidate and how they talk about the political landscape. Plus, it serves to demonstrate the data science techniques I’m learning for as a portfolio project.

If you want to play with the model yourself, I also put together an interactive javascript page for you: you can test your judgment compared to its predictions, browse the most Trumpish/Clintonish sentences and terms, and enter your own text for the model to evaluate.

To read about how the model works, I wrote a rundown with both technical and non-technical details below the tables and graphs. But without further ado, the results:

# The Trump-iest and Clinton-est Sentences and Phrases from the 2016 Campaign:

Clinton Trump
Top sentence: “That’s why the slogan of my campaign is stronger together because I think if we work together and overcome the divisiveness that sometimes sets americans against one another and instead we make some big goals and I’ve set forth some big goals, getting the economy to work for everyone, not just those at the top, making sure we have the best education system from preschool through college and making it affordable and somp[sic] else.” — Presidential Candidates Debate

Predicted Clinton: 0.99999999999
Predicted Trump: 1.04761466567e-11

Frustratingly, I couldn’t download or embed the C-SPAN video for this clip, so here are two of the other top 5 Clinton-iest sentences:

Presidential Candidate Hillary Clinton Rally in Orangeburg, South Carolina

Presidential Candidate Hillary Clinton Economic Policy Address

Top sentence: “As you know, we have done very well with the evangelicals and with religion generally speaking, if you look at what’s happened with all of the races, whether it’s in south carolina, i went there and it was supposed to be strong evangelical, and i was not supposed to win and i won in a landslide, and so many other places where you had the evangelicals and you had the heavy christian groups and it was just — it’s been an amazing journey to have — i think we won 37 different states.” — Faith and Freedom Coalition Conference

Predicted Clinton: 4.29818403092e-11
Predicted Trump: 0.999999999957

Frustratingly, I couldn’t download or embed the C-SPAN video for this clip either, so here are two of the other top 5 Trump-iest sentences:

Presidential Candidate Donald Trump Rally in Arizona

Presidential Candidate Donald Trump New York Primary Night Speech

## Top Terms

Term Multiplier
my husband 12.95
recession 10.28
attention 9.72
wall street 9.44
grateful 9.23
or us 8.39
citizens united 7.97
mother 7.20
something else 7.17
strategy 7.05
clear 6.81
kids 6.74
gun 6.69
i remember 6.51
corporations 6.51
learning 6.36
democratic 6.28
clean energy 6.24
well we 6.14
insurance 6.14
grandmother 6.12
experiences 6.00
progress 5.94
auto 5.90
climate 5.89
over again 5.85
often 5.80
a raise 5.71
immigration reform 5.62
Term Multiplier
tremendous 14.57
guy 10.25
media 8.60
does it 8.24
hillary 8.15
politicians 8.00
almost 7.83
incredible 7.42
illegal 7.16
general 7.03
frankly 6.97
border 6.89
establishment 6.84
jeb 6.76
allowed 6.72
obama 6.48
poll 6.24
by the way 6.21
bernie 6.20
ivanka 6.09
japan 5.98
politician 5.96
nice 5.93
conservative 5.90
islamic 5.77
hispanics 5.76
deals 5.47
win 5.43
guys 5.34
believe me 5.32

## Cherrypicked pairs of terms:

Clinton Trump
Term Multiplier Term Multiplier
president obama 3.27 obama 6.49
immigrants 3.40 illegal immigrants 4.87
clean energy 6.24 energy 1.97
the wealthy 4.21 wealth 2.11
learning 6.36 earning 1.38
muslims 3.46 the muslims 1.75
senator sanders 3.18 bernie 6.20

# How the Model Works:

### Defining the problem: What makes a sentence “Trump-y?”

I decided that the best way to quantify ‘Trump-iness’ of a sentence was to train a model to predict whether a given sentence was said by Trump or Clinton. The Trumpiest sentence will be the one that the predictive model would analyze and say “Yup, the chance this was Trump rather than Clinton is 99.99%”.

Along the way, with the right model, we can ‘look under the hood’ to see what factors into the decision.

Technical details:

The goal is to build a classifier that can distinguish between the candidate’s sentences optimizing for ROC_AUC, and allows us to extract meaningful/explainable coefficients.

### Gathering and processing the data:

In order to train the model, I needed large bodies of text from each candidate. I ended up scraping transcripts from events on C-SPAN.org. Unfortunately, they’re uncorrected closed caption transcripts and contained plenty of typos and misattributions. On the other hand, they’re free.

I did a bit to clean up some recurring problems like the transcript starting every quote section with “Sec. Clinton:” or including descriptions like [APPLAUSE] or [MUSIC]. (Unfortunately, they don’t reliably mark the end of the music, and C-SPAN sometimes claims that Donald Trump is the one singing ‘You Can’t Always Get What You Want.’)

Technical details:

I ended up learning to use Python’s Beautiful Soup library to identify the list of videos C-SPAN considers campaign events by the candidates, find their transcripts, and grab only the parts they supposedly said. I learned to use some basic regular expressions to do the cleaning.

My scraping tool is up on github, and is actually configured to be able to grab transcripts for other people as well.

### Converting the data into usable features

After separating the large blocks of text into sentences and then words, I had some decisions to make. In an effort to focus on interesting and meaningful content, I removed sentences that were too short or too long – “Thank you” comes up over and over, and the longest sentences tended to be errors in the transcription service. It’s a judgement call, but I wanted to keep half the sentences, which set cutoffs at 9 words and 150 words. 34,108 sentences remained.

A common technique in natural language processing is to remove the “stopwords” – common non-substantive words like articles (a, the), pronouns (you, we), and conjunctions (and, but). However, following James Pennebaker’s research, which found these words are surprisingly useful in predicting personality, I left them in.

Now we have what we need: sequences of words that the model can consider evidence of Trump-iness.

Technical details:

I used NLTK to tokenize the text into sentences, but wrote my own regular expressions to tokenize the words. I considered it important to keep contractions together and include single-character tokens, which the standard NLTK function wouldn’t have done.

I used a CountVectorizer from sklearn to extract ngrams and later selected the most important terms using a SelectFromModel with a Lasso Logistic Regression. It was a balance – more terms would typically improve accuracy, but water down the meaningfulness of each coefficient.

I tested using various additional features, like parts of speech and lemmas (using the fantastic Spacy library) and sentiment analysis (using the Textblob library) but found that they only provided marginal benefit and made the model much slower. Even just using 1-3 ngrams, I got 0.92 ROC_AUC.

### Choosing & Training the Model

One of the most interesting challenges was avoiding overfitting. Without taking countermeasures, the model could look at a typo-riddled sentence like “Wev justv don’tv winv anymorev.” and say “Aha! Every single one of those words are unique to Donald Trump, therefore this is the most Trump-like sentence ever!”

I addressed this problem in two ways: the first is by using regularization, a standard machine learning technique that penalizes a model for using larger coefficients. As a result, the model is discouraged from caring about words like ‘justv’ which might only occur two times, since they would only help identify those couple sentences. On the other hand, a word like ‘frankly’ helps identify many, many sentences and is worth taking a larger penalty to give it more importance in the model.

The other technique was to use batch predictions – dividing the sentences into 20 chunks, and evaluating each chunk by only training on the other 19. This way, if the word ‘winv’ only appears in a single chunk, the model won’t see it in the training sentences and won’t be swayed. Only words that appear throughout the campaign have a significant impact in the model.

Technical details:

The model uses a logistic regression classifier because it produces very explainable coefficients. If that weren’t a factor, I might have tried a neural net or SVM (I wouldn’t expect a random forest to do well with such sparse data.) In order to set the regularization parameters for both the final classifier and for the feature-selection Lasso Logistic Regressor, I used sklearn’s cross-validated gridsearch object, optimizing for ROC_AUC.

During the prediction process, I used a stratified Kfold to divide the data in order to ensure each chunk would have the appropriate mix of Trump and Clinton sentences. It was tempting to treat the sentences more like a time series and only use past data in the predictions, but we want to consider how similar old sentences are to the whole corpus.

### Interpreting and Visualizing the Results:

The model produced two interesting types of data: how likely the model thought each sentence was spoken by Trump or Clinton (how ‘Trumpish’ vs. ‘Clintonish’ it is), and how any particular term impacts those predicted odds. So if a sentence is predicted to be spoken by Trump with estimated 99.99% probability, the model considers it extremely Trumpish.

The term’s multipliers indicate how each word or phrase impacts the predicted odds. The model starts at 1:1 (50%/50%), and let’s say the sentence includes the word “incredible” – a Trump multiplier of 7.42. The odds are now 7.42 : 1, or roughly 88% in favor of Trump. If the model then sees the word “grandmother” – a Clinton multiplier of 6.12 – its estimated odds become 7.42 : 6.12, (or 1.12 : 1), roughly 55% Trump. Each term has a multiplying effect, so a 4x word and 2x word together have as much impact as an 8x word – not 6x.

Technical details:

In order to visualize the results, I spent a bunch of time tweaking the matplotlib package to generate a graph of coefficients, which I used for the pronouns above. I made sure to use a logarithmic scale, since the terms are multiplicative.

In addition, I decided to teach myself enough javascript to learn to use the D3 library – allowing interactive visualizations and the guessing game where players can try to figure out who said a given random sentence from the campaign trail. There are a lot of ways the code could be improved, but I’m pleased with how it turned out given that I didn’t know any D3 prior to this project.

## An Atheist’s Defense of Rituals: Ceremonies as Traffic Lights

The idea of a coming-of-age ceremony has always been a bit strange to me as an atheist. Sure, I attended more than my fair share of Bat and Bar Mitzvahs in middle school. But it always struck me as odd for us to pretend that someone “became an adult” on a particular day, rather than acknowledging it was a gradual process of maturation over time. Why can’t we just all treat people as their maturity level deserves?

The same goes with weddings – does a couple’s relationship really change in a significant way marked by a ceremony? Or do two people gradually fall in love and grow committed to each other over time? Moving in with each other marks a discrete change, but what does “married” change about the relationship?

But my thinking has been evolving since reading this fantastic post about rituals by Brett and Kate McKay at The Art of Manliness. Not only do the rituals acknowledge a change, they use psychological and social reinforcement to help the individuals make the transition more fully:

One of the primary functions of ritual is to redefine personal and social identity and move individuals from one status to another: boy to man, single to married, childless to parent, life to death, and so on.

Left to follow their natural course, transitions often become murky, awkward, and protracted. Many life transitions come with certain privileges and responsibilities, but without a ritual that clearly bestows a new status, you feel unsure of when to assume the new role. When you simply slide from one stage of your life into another, you can end up feeling between worlds – not quite one thing but not quite another. This fuzzy state creates a kind of limbo often marked by a lack of motivation and direction; since you don’t know where you are on the map, you don’t know which way to start heading.

Just thinking your way to a new status isn’t very effective: “Okay, now I’m a man.” The thought just pings around inside your head and feels inherently unreal. Rituals provide an outward manifestation of an inner change, and in so doing help make life’s transitions and transformations more tangible and psychologically resonant.

Brett and Kate McKay cover a range of aspects of rituals, but I was particular struck by the game theory implications of these ceremonies. By coordinating society’s expectations in a very public manner, transition rituals act like traffic lights to make people feel comfortable and confident in their course of action.

### The Value of Traffic Lights

Traffic lights are a common example in game theory. Imagine that you’re driving toward an unmarked intersection and see another car approaching from the right. You’re faced with a decision: do you keep going, or brake to a stop?

If you assume they’re going to keep driving, you want to stop and let them pass. If you’re wrong, you both lose time and there’s an awkward pause while you signal to each other to go.

If you assume they’re going to stop, you get to keep going and maintain your speed. Of course, if you’re wrong and they keep barreling forward, you risk a deadly accident.

Things go much more smoothly when there are clear street signs or, better yet, a traffic light coordinating everyone’s expectations.

### Ceremonies as Traffic Lights

Now, misjudging a teenager’s maturity is unlikely to result in a deadly accident. But, with reduced stakes, the model still applies.

As a teen gets older, members of society don’t always know how to treat him – as a kid or adult. Each type of misaligned expectations is a different failure mode: If you treat him as a kid when he expected to be treated as an adult, he might feel resentful of the “overbearing adult”. If you treat him as an adult when he was expecting to be treated as a kid, he might not take responsibility for himself.

A coming-of-age ritual acts like the traffic light to minimize those failure modes. At a Bar or Bat Mitzvah, members of society gather with the teenager and essentially publicly signal “Ok everyone, we’re switching our expectations… wait for it… Now!”

It’s important that the information is known by all to be known to all – what Steven Pinker calls common or mutual knowledge:

“In common knowledge, not only does A know x and B know x, but A knows that B knows x, and B knows that A knows x, and A knows that B knows that A knows x, ad infinitum.”

If you weren’t sure that the oncoming car could see their traffic light, it would be almost as bad as if there were no light at all. You couldn’t trust your green light because they might not stop. Not only do you need to know your role, but you need to know that everyone knows their role and trusts that you know yours… etc.

Public ceremonies gather everyone to one place, creating that common knowledge. The teenager knows that everyone expects him to act as an adult, society knows that he expects them to treat him as one, and everyone knows that those expectations are shared. Equipped with this knowledge, the teen can count on consistent social reinforcement to minimize awkwardness and help him adopt his new identity.

Obviously, these rituals are imperfect – Along with the socially-defined parts of identity, there are internal factors that make someone more or less ready to be an adult. Quite frankly, setting 13 as the age of adulthood is probably too young.

But that just means we should tweak the rituals to better fit our modern world. After all, we have precise engineering to set traffic light schedules, and it still doesn’t seem perfect (this XKCD comes to mind).

That’s what makes society and civilization powerful. We’re social creatures, and feel better when we feel comfortable in our identity – either as a child or adult, as single or married, as grieving or ready to move on. Transition rituals serve an important and powerful role in coordinating those identities.

We shouldn’t necessarily respect them blindly, but I definitely respect society’s rituals more after thinking this through.

To take an excerpt from a poem by Bruce Hawkins:

Three in the morning, Dad, good citizen
stopped, waited, looked left, right.
He had been driving nine hundred miles,
had nearly a hundred more to go,
but if there was any impatience
it was only the steady growl of the engine
which could just as easily be called a purr.

I chided him for stopping;
he told me our civilization is founded
on people stopping for lights at three in the morning.

## The Matrix Meets Braid: Artificial Brains in Gunfights

It’s The Matrix meets Braid: a first-person shooter video game “where the time moves only when you move.” You can stare at the bullets streaking toward you as long as you like, but moving to dodge them causes the enemies and bullets to move forward in time as well.

The game is called SUPERHOT, and the designers describe it by saying “With this simple mechanic we’ve been able to create gameplay that’s not all about reflexes – the player’s main weapon is careful aiming and smart planning – while not compromising on the dynamic feeling of the game.”

Here’s the trailer:

I’ve always loved questions about what it would be like to distort time for yourself relative to the rest of the universe (and the potential unintended consequences, as we explored in discussing why The Flash is in a special hell.)

In Superhot, it’s not that you can distort time exactly – after all, whenever you take a step, your enemies get the same amount of time to take a step themselves. Instead, your brain is running as fast as it likes while (the rest of) your body remains in the same time stream as everything else.

And then it struck me: this might be close to the experience of an emulated brain housed in a regular-sized body.

Let’s say that, in the future, we artificially replicate/emulate human minds on computers. And let’s put an emulated human mind inside a physical, robotic body. The limits on how fast it can think are its hardware and its programming. As technology and processor speeds improve, the “person” could think faster and faster and would experience the outside world as moving slower and slower in comparison.

… but even though you might have a ridiculously high processing speed to think and analyze a situation, your physical body is still bound by the normal laws of physics. Moving your arms or legs requires moving forward in the same stream of time as everyone else. In order to, say, turn your head to look to your left and gather more information, you need to let time pass for your enemies, too.

Robin Hanson, professor of economics at George Mason University and author of Overcoming Bias, has put a lot of thought into the implications of whole-brain emulation. So I asked him:

Is Superhot what an emulated human would experience in a gunfight?

An em could usually speed up its mind to deal with critical situations, though this would cost more per objective second. So a first-person shooter where time only moves when you do does move in the direction of letting the gamer experience action in an em world. Even better would be to let the gamer change the rate at which game-time seems to move, to have a limited gamer-time budget to spend, and to give other non-human game characters a similar ability.”

He’s right: thinking faster would require running more cycles per second, which takes resources. And yeah, you would need infinite processing speed to think indefinitely while the rest of the world was frozen. It would be more consistent to add a “mental cycle” budget that ran down at a constant rate from the gamer’s external point of view.

I don’t know about you, but I would buy that game! (Even if a multi-player mode would be impossible.)

## Why Decision Theory Tells You to Eat ALL the Cupcakes

Imagine that you have a big task coming up that requires an unknown amount of willpower – you might have enough willpower to finish, you might not. You’re gearing up to start when suddenly you see a delicious-looking cupcake on the table. Do you indulge in eating it? According to psychology research and decision-theory models, the answer isn’t simple.

If you resist the temptation to eat the cupcake, current research indicates that you’ve depleted your stores of willpower (psychologists call it ego depletion), which causes you to be less likely to have the willpower to finish your big task. So maybe you should save your willpower for the big task ahead and eat it!

…But if you’re convinced already, hold on a second. How easily you give in to temptation gives evidence about your underlying strength of will. After all, someone with weak willpower will find the reasons to indulge more persuasive. If you end up succumbing to the temptation, it’s evidence that you’re a person with weaker willpower, and are thus less likely to finish your big task.

How can eating the cupcake cause you to be more likely to succeed while also giving evidence that you’re more likely to fail?

### Conflicting Decision Theory Models

The strangeness lies in the difference between two conflicting models of how to make decisions. Luke Muehlhauser describes them well in his Decision Theory FAQ:

This is not a “merely verbal” dispute (Chalmers 2011). Decision theorists have offered different algorithms for making a choice, and they have different outcomes. Translated into English, the [second] algorithm (evidential decision theory or EDT) says “Take actions such that you would be glad to receive the news that you had taken them.” The [first] algorithm (causal decision theory or CDT) says “Take actions which you expect to have a positive effect on the world.”

The crux of the matter is how to handle the fact that we don’t know how much underlying willpower we started with.

Causal Decision Theory asks, “How can you cause yourself to have the most willpower?”

It focuses on the fact that, in any state, spending willpower resisting the cupcake causes ego depletion. Because of that, it says our underlying amount of willpower is irrelevant to the decision. The recommendation stays the same regardless: eat the cupcake.

Evidential Decision Theory asks, “What will give evidence that you’re likely to have a lot of willpower?”

We don’t know whether we’re starting with strong or weak will, but our actions can reveal that one state or another is more likely. It’s not that we can change the past – Evidential Decision Theory doesn’t look for that causal link – but our choice indicates which possible version of the past we came from.

Yes, seeing someone undergo ego depletion would be evidence that they lost a bit of willpower.  But watching them resist the cupcake would probably be much stronger evidence that they have plenty to spare.  So you would rather “receive news” that you had resisted the cupcake.

### A Third Option

Each of these models has strengths and weaknesses, and a number of thought experiments – especially the famous Newcomb’s Paradox – have sparked ongoing discussions and disagreements about what decision theory model is best.

One attempt to improve on standard models is Timeless Decision Theory, a method devised by Eliezer Yudkowsky of the Machine Intelligence Research Institute.  Alex Altair recently wrote up an overview, stating in the paper’s abstract:

When formulated using Bayesian networks, two standard decision algorithms (Evidential Decision Theory and Causal Decision Theory) can be shown to fail systematically when faced with aspects of the prisoner’s dilemma and so-called “Newcomblike” problems. We describe a new form of decision algorithm, called Timeless Decision Theory, which consistently wins on these problems.

It sounds promising, and I can’t wait to read it.

### But Back to the Cupcakes

For our particular cupcake dilemma, there’s a way out:

Precommit. You need to promise – right now! – to always eat the cupcake when it’s presented to you. That way you don’t spend any willpower on resisting temptation, but your indulgence doesn’t give any evidence of a weak underlying will.

And that, ladies and gentlemen, is my new favorite excuse for why I ate all the cupcakes.

## How has Bayes’ Rule changed the way I think?

People talk about how Bayes’ Rule is so central to rationality, and I agree. But given that I don’t go around plugging numbers into the equation in my daily life, how does Bayes actually affect my thinking?
A short answer, in my new video below:

(This is basically what the title of this blog was meant to convey — quantifying your uncertainty.)