Review: Algorithms to Live By, by Brian Christian & Tom Griffiths

[Russ Allbery]

Algorithms to Live By
by Brian Christian & Tom Griffiths

Publisher: Henry Holt and Company
Copyright: April 2016
ISBN:      1-62779-037-3
Format:    Kindle
Pages:     255

Another read for the work book club. This was my favorite to date,
apart from the books I recommended myself.

One of the foundations of computer science as a field of study is
research into algorithms: how do we solve problems efficiently using
computer programs? This is a largely mathematical field, but it's often
less about ideal or theoretical solutions and more about making the
most efficient use of limited resources and arriving at an adequate, if
not perfect, answer. Many of these problems are either day-to-day human
problems or are closely related to them; after all, the purpose of
computer science is to solve practical problems with computers. The
question asked by Algorithms to Live By is "can we reverse this?": can
we learn lessons from computer science's approach to problems that
would help us make day-to-day decisions?

There's a lot of interesting material in the eleven chapters of this
book, but there's also an amusing theme: humans are already very good
at this. Many chapters start with an examination of algorithms and
mathematical analysis of problems, dive into a discussion of how we can
use those results to make better decisions, then talks about studies of
the decisions humans actually make... and discovers that humans are
already applying ad hoc versions of the best algorithms we've come up
with, given the constraints of typical life situations. It tends to
undermine the stated goal of the book. Thankfully, it in no way
undermines interesting discussion of general classes of problems, how
computer science has tackled them, and what we've learned about the
mathematical and technical shapes of those problems. There's a bit less
self-help utility here than I think the authors had intended, but lots
of food for thought.

(That said, it's worth considering whether this congruence is less
because humans are already good at this and more because our algorithms
are designed from human intuition. Maybe our best algorithms just
reflect human thinking. In some cases we've checked our solutions
against mathematical ideals, but in other cases they're still just our
best guesses to date.)

This is the sort of a book where a chapter listing is an important part
of the review. The areas of algorithms discussed here are optimal
stopping, explore/exploit decisions (when to go with the best thing
you've found and when to look for something better), sorting, caching,
scheduling, Bayes's rule (and prediction in general), overfitting when
building models, relaxation (solving an easier problem than your actual
problem), randomized algorithms, a collection of networking algorithms,
and finally game theory. Each of these has useful insights and
thought-provoking discussion of how these sometimes-theoretical
concepts map surprisingly well onto daily problems. The book concludes
with a discussion of "computational kindness": an encouragement to
reduce the required computation and complexity penalty for both
yourself and the people you interact with.

If you have a computer science background (as I do), many of these will
be familiar concepts, and you might be dubious that a popularization
would tell you much that's new. Give this book a shot, though; the
analogies are less stretched than you might fear, and the authors are
both careful and smart about how they apply these principles. This book
passes with flying colors a key sanity check: the chapters on topics
that I know well or have thought about a lot make few or no obvious
errors and say useful and important things. For example, the scheduling
chapter, which unsurprisingly is about time management, surpasses more
than half of the time management literature by jumping straight to the
heart of most time management problems: if you're going to do
everything on a list, it rarely matters the order in which you do it,
so the hardest scheduling problems are about deciding what not to do
rather than deciding order.

The point in the book where the authors won my heart completely was in
the chapter on Bayes's rule. Much of the chapter is about Bayesian
priors, and how one's knowledge of past events is a vital part of
analysis of future probabilities. The authors then discuss the
(in)famous marshmallow experiment, in which children are given one
marshmallow and told that if they refrain from eating it until the
researcher returns, they'll get two marshmallows. Refraining from
eating the marshmallow (delayed gratification, in the psychological
literature) was found to be associated with better life outcomes years
down the road. This experiment has been used and abused for years for
all sorts of propaganda about how trading immediate pleasure for future
gains leads to a successful life, and how failure in life is because of
inability to delay gratification. More evil analyses have (of course)
tied that capability to ethnicity, with predictably racist results.

I have kind of a thing about the marshmallow experiment. It's a topic
that reliably sends me off into angry rants.

Algorithms to Live By is the only book I have ever read to mention the
marshmallow experiment and then apply the analysis that I find far more
convincing. This is not a test of innate capability in the children;
it's a test of their Bayesian priors. When does it make perfect sense
to eat the marshmallow immediately instead of waiting for a reward?
When their past experience tells them that adults are unreliable, can't
be trusted, disappear for unpredictable lengths of time, and lie. And,
even better, the authors supported this analysis with both a follow-up
study I hadn't heard of before and with the observation that some
children would wait for some time and then "give in." This makes
perfect sense if they were subconsciously using a Bayesian model with
poor priors.

This is a great book. It may try a bit too hard in places
(applicability of the math of optimal stopping to everyday life is more
contingent and strained than I think the authors want to admit), and
some of this will be familiar if you've studied algorithms. But the
writing is clear, succinct, and very well-edited. No part of the book
outlives its welcome; the discussion moves right along. If you find
yourself going "I know all this already," you'll still probably
encounter a new concept or neat explanation in a few more pages. And
sometimes the authors make connections that never would have occurred
to me but feel right in retrospect, such as relating exponential
backoff in networking protocols to choosing punishments in the criminal
justice system. Or the realization that our modern communication world
is not constantly connected, it's constantly buffered, and many of us
are suffering from the characteristic signs of buffer bloat.

I don't think you have to be a CS major, or know much about math, to
read this book. There is a lot of mathematical details in the end notes
if you want to dive in, but the main text is almost always readable and
clear, at least so far as I could tell (as someone who was a CS major
and has taken a lot of math, so a grain of salt may be indicated). And
it still has a lot to offer even if you've studied algorithms for
years.

The more I read of this book, the more I liked it. Definitely
recommended if you like reading this sort of analysis of life.

Rating: 9 out of 10

Reviewed: 2017-10-22

-- 
Russ Allbery (eagle at eyrie.org)              <http://www.eyrie.org/~eagle/>