The suboptimal life.
It's ok if we're not optimizing anything.
Programming Note: I’m traveling this week and the posting schedule will be erratic.
Continuing down my path of optimization apostasy, I wanted to respond to Sarah Dean and Max Raginsky about the troubles of Inverse Optimization. If everything is optimization, then everything we see before us must be optimizing something. Inverse Optimization aims to determine which optimization problem gives rise to observed dynamical behavior.
The problem is that anything can be optimal for something. I can imagine a variety of functions for which my coffee cup is a maximum. But why is it helpful to think about what observed reality optimizes? Perhaps it could inform the design of future coffee cups. It could give insights into how my coffee cup came to be in the first place. But for Inverse Optimization to be useful, the fit optimization problem must extrapolate to explain new observations.
Physics provides many compelling examples of the power of Inverse Optimization. Light takes paths that minimize travel times (Fermat’s Principle). Much of modern physics can be derived through principles of least action. Of all possible realities, ours is the one that makes stationary some Lagrangian. How generally can we look at non-engineered systems and determine what they optimize?
I worry that as we push away from the elegantly constrained world of physics, the Inverse Optimization mindset is more harmful than helpful. For example, “Everything Is Optimization” feeds a purely adaptationist view of evolution. For the adaptionist, traits exist because they improve the success of reproduction. But the dynamics of evolution are far more complex. Evolutionary biologists have argued strongly against these adaptionist views, most famously in the essay “The Spandrels of San Marco and the Panglossian Paradigm” by Gould and Lewontin. Lewontin has written extensively about why evolutionary selection can’t entirely be explained by optimizing.
Another prime example where Inverse Optimization fails is (all of) economics. Microeconomics tells us that people maximize utilities and we just have to apply Inverse Optimization to figure out what these utilities are. But countless data show people don’t maximize. And we can’t correct for this lack of maximization with psychological babble about predictable irrationality. Economics and behavioral economics is neither predictive nor generalizable. Fitting an optimization model in one context doesn’t generalize to another context. The evidentiary record of Game Theory shows people play all sorts of different ways depending on culture and experimental set-up, and it is impossible to predict from one experiment how any other will go. What’s worse is that the behavioral and evolutionary econometric “fixes” have all been shown to be unreproducible storytelling or, in recent years, outright fraud.
Making matters worse, when you combine the adaptionist mindset with game theory, you convince yourself that all dominant power structures are explained by strategic just-so stories. Inverse Optimization leads to grotesque ends like cultural evolutionary theory and effective altruism. That somehow we can back out what is optimal from our experience and then think that we can reverse engineer ethics and morality as optimal. This worldview is simultaneously facile and authoritarian yet embraced by many populist intellectuals in the name of science.
Am I arguing there’s a slippery slope to the optimization mindset? I worry that I might be.
I am in total agreement with all of this! My personal view is that (almost) everything is control, not optimization, but control understood in the Willemsian sense: restricting the behavior of one system by interconnecting it with another. There need not be any notion of optimality here: some system trajectories are simply forbidden, and then we can talk about how (or why) some of the allowed trajectories are selected or actualized. Borrowing a pithy phrase from Peter Gould (no relation to Stephen Jay Gould), control in general is about "allowing, forbidding, but not requiring." So, for example, nonadaptationist evolutionary theories, such as genetic drift, fall into this category. The environment simply rejects (or forbids) nonviable options, but which particular path is taken depends on lots of factors, including local optimization-like mechanisms without any global optimization. This is also the case in social/economic settings: Laws, regulations, customs, etc. forbid certain outcomes, but which of the allowed outcomes is selected may not be the result of optimization, although norms, values, etc. certainly come into play.
A pet peeve of mine is the muddied thinking that arises from conflating biological phenomena and computer algorithms. On the computer science side, it obscures the fact that we are responsible for the outcomes since it's our choices that determine the system's behavior. And on the biology side, we collapse the complexity into the crude and comparatively simplistic world of computation. So I am looking forward to reading the linked critique of adaptionist views! Many of the points you make here ring true.
Now let me expand on the "we are responsible" point. Where "everything is optimization" is most meaningful for me is when "everything" refers to "things that we build". To reiterate my comment on your last post, I think that automated systems are most clearly understood through the optimization lens. For example, the public "For you" Twitter algorithm ranks tweets as a weighted sum of the probability of various types of engagement actions: https://slides.com/sarahdean-2/deck-4f230b?token=HmEsTam3#/7
It is fun to try to interpret these weights. Why is the weight on "report" -369 rather than -370? Why is the weight on "watch half of the video" 0.005 when all other weights are O(1)? I can only speculate that upranking videos led to a bad user experience but there was internal pressure to "promote videos" so this was the compromise. Or maybe the predicted probability is often very close to 1 (since videos autoplay?) and 0.005 brings them into the same scale as other predicted engagement actions.
All this to say, these weights were almost certainly decided with reference to some metrics, KPIs, etc. Many people have made the point that open sourcing the algorithms is meaningless without access to the engagement models. I would argue that these metrics/KPIs are similarly important---from the perspective of critique and from the perspective of thinking clearly about platform design. I agree with something Neal Parikh said on Twitter: there is clarifying value in a declarative representation of what a system is built for.