Chapter 12: Satisficing

Why "Good Enough" Beats Optimal in Almost Every Real System

"The capacity of the human mind for formulating and solving complex problems is very small compared with the size of the problems whose solution is required for objectively rational behavior in the real world." -- Herbert A. Simon, Models of Man (1957)

12.1 The Cereal Aisle

You are standing in a grocery store, staring at the cereal aisle. There are, depending on the store, between eighty and three hundred varieties of breakfast cereal arranged in a wall of color and competing claims. High fiber, low sugar, whole grain, gluten-free, organic, fortified, kid-friendly, heart-healthy, keto-compatible, ancient grains, with freeze-dried strawberries, without artificial flavors, in a resealable bag, in a box with a puzzle on the back.

The economics textbook says you should choose the box that maximizes your utility. To do this, you would need to assess your preferences across every relevant dimension -- taste, nutrition, price, brand loyalty, ingredient quality, environmental footprint of the packaging, the probability that your children will actually eat it -- assign weights to each dimension, compute a utility score for each of the three hundred options, and select the one with the highest score.

Nobody does this. Nobody has ever done this. Nobody could do this, not even with unlimited time and a spreadsheet, because many of the relevant dimensions are incommensurable (how do you trade off taste against environmental impact?), many of the necessary inputs are unknown (how much will you enjoy a cereal you have never tried?), and the computational cost of the full optimization dwarfs the value of the decision. The difference in life satisfaction between the "optimal" cereal and a merely good one is vanishingly small. The time spent computing the optimum is time you will never get back.

So what do you actually do? You walk down the aisle, glance at a few familiar options, maybe notice something new, pick one that seems fine, and move on. You do not maximize. You satisfice -- you search until you find an option that meets your requirements, and then you stop searching.

This is not laziness. This is not irrationality. This is, according to one of the most important insights in the history of the social sciences, the only rational response to the actual conditions under which decisions must be made.

The man who saw this most clearly was Herbert Alexander Simon, and his insight -- that the classical model of rational choice is not merely unrealistic but fundamentally wrong, and that "good enough" is not a compromise but often the genuinely optimal strategy -- is the subject of this chapter.

Fast Track: Satisficing means searching for an option that meets a threshold of acceptability rather than computing the theoretically best option. Herbert Simon coined the term to describe how people actually make decisions under real-world constraints of limited time, information, and computational capacity. This chapter shows that satisficing is not a failure of rationality but an adaptive strategy that appears independently in evolution, engineering, military planning, chess, and everyday life -- and that it often outperforms theoretical optimization.

Deep Dive: The implications of satisficing run deeper than decision strategy. Simon's bounded rationality challenges the foundational assumptions of neoclassical economics, rational choice theory, and any framework that models agents as optimizers. It connects to computational complexity theory (some optimization problems are provably intractable), to ecological rationality (the fit between a strategy and its environment), and to the paradox of choice (why more options can make us worse off). The chapter also previews how excessive optimization leads to overfitting (Ch. 14) and how optimizing the wrong metric leads to Goodhart's law (Ch. 15).

12.2 The Man Who Saw Through Optimization

Herbert Simon was born in Milwaukee in 1916 and went on to become one of the most intellectually wide-ranging thinkers of the twentieth century. He won the Nobel Prize in Economics in 1978, the Turing Award in computer science in 1975, and made foundational contributions to political science, cognitive psychology, artificial intelligence, and organizational theory. He is one of a very small number of people who made lasting contributions to more than four academic disciplines -- a genuine polymath in an era of increasing specialization.

Simon's central insight emerged from observing how decisions are actually made inside organizations. Classical economics assumed that firms maximize profit. Classical decision theory assumed that rational agents maximize expected utility. Simon looked at real organizations -- city governments, corporations, military bureaucracies -- and saw something entirely different. He saw administrators making decisions with fragmentary information, under time pressure, with cognitive limitations that made anything resembling full optimization impossible.

From this observation came the concept of bounded rationality: the idea that human rationality is limited -- bounded -- by the information available, the cognitive capacity of the mind, and the time available for decision-making. Bounded rationality does not mean irrationality. It means rationality within limits. Given those limits, the rational strategy is not to optimize (which is impossible) but to satisfice (which is achievable and, as it turns out, frequently superior).

Simon coined the word "satisfice" by combining "satisfy" and "suffice." A satisficing strategy works like this: define a threshold of acceptability (an aspiration level), search through available options, and select the first option that meets or exceeds the threshold. If no option meets the threshold, lower the threshold and search again. If many options meet the threshold easily, raise the threshold for next time.

This sounds simple, almost trivially so. Its profundity lies in what it implies about the nature of rationality itself. If optimization is impossible in most real-world situations -- and Simon argued persuasively that it is -- then the textbook model of rational choice is not a description of ideal behavior that real agents approximate imperfectly. It is a fantasy, a mathematical convenience that misrepresents the actual structure of rational decision-making. Satisficing is not a poor substitute for optimization. It is the real thing. Optimization is the illusion.

Connection to Chapter 7 (Gradient Descent): Simon's satisficing has a deep structural parallel to the local optimization problem in gradient descent. Gradient descent does not find the global optimum -- it finds a local one and stops. This is not a failure of the algorithm; for many real problems, a good local optimum is all you need, and the cost of searching for the global optimum is prohibitive. Satisficing is the decision-theoretic equivalent of accepting a good local minimum rather than exhaustively searching for the global one.

12.3 Evolution Does Not Optimize

If there is one system that ought to find the best possible solution, surely it is natural selection. Evolution has had billions of years, trillions of organisms, and the harshest fitness criterion imaginable -- survive and reproduce, or your lineage ends forever. If any process could optimize, evolution should be it.

Evolution does not optimize. Evolution satisfices.

Consider the human eye. It is a remarkable organ -- capable of detecting a single photon, resolving fine detail, adapting to light levels spanning ten orders of magnitude, and processing visual information with stunning speed. It is also, from an engineering perspective, built backwards. The photoreceptors in the vertebrate retina point away from the incoming light, toward the back of the eye. The nerve fibers and blood vessels lie in front of the receptors, partially blocking the light. The nerve fibers converge into the optic nerve, which exits through the retina, creating a blind spot in each eye. The cephalopod eye -- found in octopuses and squid -- has the photoreceptors pointing forward, toward the light, with no blind spot. The octopus eye is, by any engineering standard, better designed.

Why does the vertebrate eye have its retina backwards? Because that is what happened to work during the early evolution of the vertebrate body plan, hundreds of millions of years ago, and evolution has no mechanism for starting over. It cannot redesign from scratch. It can only modify what already exists. Each generation is constrained by the developmental architecture inherited from the previous generation. Evolution does not search the space of all possible eyes and select the best one. It searches the narrow space of modifications to the existing eye that happen to arise through random mutation, and it selects any modification that is good enough to survive and reproduce.

"Good enough to survive" is the only standard evolution applies. This is satisficing in its purest form. There is no aspiration toward perfection, no comparison against an ideal design, no mechanism for evaluating whether a better solution exists elsewhere in the design space. There is only the binary filter: does this organism survive long enough to reproduce? If yes, its design persists. If no, it vanishes. The result is a biosphere full of organisms that are exquisitely adequate -- beautifully adapted to their environments, but carrying the scars of historical accident, developmental constraint, and paths not taken.

The recurrent laryngeal nerve in the giraffe is a famous example. This nerve connects the brain to the larynx, a distance of a few inches. But because of the way the nerve was routed in the ancestral fish body plan -- looping around the aortic arch -- the giraffe's recurrent laryngeal nerve travels from the brain all the way down the neck, around the aorta near the heart, and all the way back up the neck to the larynx, a detour of approximately fifteen feet. No engineer would design this. But evolution does not engineer. It tinkers. It satisfices. The nerve works. It is good enough. That is all that matters.

Connection to Chapter 8 (Explore/Exploit): Evolution's satisficing strategy is deeply connected to the explore/exploit tradeoff. Mutation and recombination provide exploration -- random variation in the design space. Natural selection provides exploitation -- the retention of variations that work. But evolution's "exploitation" is not optimization. It is satisficing: any variant that clears the survival threshold is exploited, regardless of whether a better variant exists somewhere in the unexplored space. Evolution is the ultimate satisficer because it has no capacity for global comparison.

🔄 Check Your Understanding

1. What is the difference between satisficing and optimizing? Provide an example from your own daily life where you satisfice rather than optimize.
2. In what sense is the backwards vertebrate retina evidence for satisficing rather than optimization in evolution?
3. How does Simon's concept of bounded rationality challenge the standard economic model of rational choice?

12.4 Chess Engines and the Impossibility of Perfection

If evolution cannot optimize because it lacks foresight, perhaps a system with explicit intelligence and computational power can do better. Consider chess. Chess is a game of perfect information -- both players can see the entire board. It has fixed rules, a finite (though astronomically large) number of possible positions, and a clear objective: checkmate the opponent's king. If optimization is possible anywhere, it should be possible in chess.

It is not. The number of possible chess games has been estimated at roughly 10 to the power of 120. This number -- known as the Shannon number, after Claude Shannon, the father of information theory (whose work we encountered in Chapter 6) -- exceeds the number of atoms in the observable universe by a factor so large that the comparison is meaningless. A computer that evaluated one position per nanosecond and had been running since the Big Bang would not have scratched the surface of the game tree.

Even Deep Blue, the IBM supercomputer that defeated world champion Garry Kasparov in 1997, did not play optimal chess. It could not. No computer can. What Deep Blue did was satisfice at an extraordinarily high level. It searched millions of positions per second using alpha-beta pruning (a technique that eliminates branches of the game tree that cannot possibly be better than already-discovered options), evaluated positions using a handcrafted evaluation function (which assessed "how good" a position looked based on material balance, piece activity, king safety, pawn structure, and other features), and selected moves that its evaluation function judged to be good enough. The evaluation function was not a measure of objective quality -- it was a heuristic, a shortcut, a way of saying "this position looks roughly this good" without calculating every possible continuation.

Modern chess engines like Stockfish and AlphaZero are far stronger than Deep Blue, but they still satisfice. Stockfish uses a more sophisticated search algorithm and a more nuanced evaluation function, but it still prunes the vast majority of possible continuations without examining them. AlphaZero, the neural-network-based engine developed by DeepMind, learned its evaluation function through self-play rather than having it hand-coded by human programmers, but it still evaluates positions heuristically and selects moves that are good enough according to its learned criteria.

The lesson is profound. Chess is one of the most constrained, well-defined problems in existence -- discrete, deterministic, fully observable, with unambiguous win/loss criteria. And even in chess, optimization is computationally intractable. The space of possibilities is simply too large. If optimization is impossible in chess, what hope does it have in domains that are continuous, stochastic, partially observable, and ill-defined -- which is to say, in virtually every real-world decision problem?

This is not a coincidence. It is a consequence of computational intractability -- the mathematical discovery, formalized in the 1970s by Stephen Cook and Richard Karp, that many natural optimization problems belong to a class (NP-hard) for which no efficient solution algorithm is known or believed to exist. The traveling salesman problem (find the shortest route visiting all cities), the protein folding problem (find the lowest-energy configuration of a protein chain), the scheduling problem (assign tasks to resources to minimize total time) -- all of these are computationally intractable in their general form. The optimal solution exists in principle but cannot be found in practice because the search space grows exponentially with the size of the problem.

Satisficing is not a concession to human cognitive limitation. It is a concession to mathematical reality. Some problems cannot be optimized, period, regardless of how much intelligence or computation you throw at them. For these problems, heuristics -- fast, frugal, good-enough strategies -- are not inferior substitutes for optimization. They are the best that any agent, natural or artificial, can do.

12.5 Military Strategy: No Plan Survives Contact

"No plan of operations extends with any certainty beyond the first contact with the main hostile force."

This observation, attributed to Helmuth von Moltke the Elder, the Prussian field marshal who masterminded the victories of the Franco-Prussian War, is one of the most famous maxims in military history. It is often paraphrased as "no plan survives contact with the enemy," and it captures a truth that extends far beyond warfare: the real world is too complex, too unpredictable, and too adversarial for any plan to be optimal in advance.

Von Moltke's insight was not that planning is useless. He was a meticulous planner. His insight was that the purpose of planning is not to produce the optimal plan but to prepare the organization to adapt when the plan inevitably fails. Planning, in Von Moltke's framework, is exploration -- a way of thinking through possibilities, building shared mental models, and identifying decision points. Execution is adaptation -- the continuous revision of the plan in light of what actually happens. The plan is the starting point, not the destination. It is satisficing applied to strategy: develop a plan that is good enough to get started, then adjust in real time.

This philosophy stands in stark contrast to the centralized, optimizing approach to military planning exemplified by the Schlieffen Plan of 1914. Count Alfred von Schlieffen, chief of the German General Staff, spent years developing a detailed plan for a two-front war against France and Russia. The plan specified troop movements, railroad schedules, and supply routes down to the day. It was, in a sense, an attempt at optimization -- a single, comprehensive plan designed to achieve victory through precise execution.

The Schlieffen Plan failed catastrophically. Not because it was poorly conceived in the abstract, but because reality did not cooperate with the plan's assumptions. Belgian resistance was fiercer than expected. British forces arrived faster than anticipated. Russian mobilization was quicker than the plan allowed for. Supply lines broke down. Communication failed. The plan, which had been optimized for a specific set of assumptions, shattered on contact with an environment that violated those assumptions.

This is the fundamental vulnerability of optimization: it is brittle. An optimized plan is optimized for a specific set of conditions. When conditions change -- and in any complex, adversarial environment, conditions always change -- the optimized plan may perform far worse than a merely good plan that was designed for robustness rather than peak performance.

Von Moltke understood this. His approach -- which modern military theorists call mission-type tactics or Auftragstaktik -- delegated decision-making to lower-level commanders who could adapt to local conditions. Rather than specifying exactly what each unit should do (which would be the "optimal" approach if the environment were predictable), he specified the objective and left the method to the commander on the ground. This is satisficing at the organizational level: do not try to compute the optimal action for every possible contingency. Instead, define what "good enough" looks like (the mission objective) and trust that competent people will find a way to achieve it.

Connection to Chapter 9 (Distributed vs. Centralized): Von Moltke's Auftragstaktik is a case study in the power of distributed decision-making, which we explored in Chapter 9. Centralized optimization requires perfect information flowing to a central authority and perfect execution flowing back -- both of which are impossible in the fog of war. Distributed satisficing requires only that each decision-maker understands the overall objective and is empowered to find a good-enough solution locally. The robustness of the distributed approach comes precisely from its refusal to optimize globally.

🔄 Check Your Understanding

1. Why is chess computationally intractable despite being a game of perfect information with fixed rules?
2. How did the failure of the Schlieffen Plan illustrate the brittleness of optimization?
3. In Von Moltke's Auftragstaktik, what takes the place of the optimal plan? How is this an example of satisficing?

12.6 Engineering Tolerances: The Cost of the Last One Percent

If you ask an engineer to build a part that is exactly 10.000 centimeters long, the engineer will ask you a question: "How exact?"

This is not a philosophical question. It is a deeply practical one. A part machined to a tolerance of plus or minus one millimeter is cheap and fast to produce. A part machined to plus or minus one hundredth of a millimeter is expensive. A part machined to plus or minus one micron is extraordinarily expensive. And a part machined to atomic precision -- to within the diameter of a single atom -- is essentially impossible with current technology, and if it were possible, it would cost more than the rest of the entire machine combined.

The relationship between precision and cost is not linear. It is exponential, or worse. Getting from 90 percent of the way to perfection to 99 percent costs roughly as much as getting from 0 to 90 percent. Getting from 99 percent to 99.9 percent costs roughly as much again. This is the engineering version of the Pareto principle -- the 80/20 rule that we encountered in our discussion of power laws (Chapter 4). The last 20 percent of quality costs 80 percent of the effort. The last 1 percent costs more than everything else combined.

Engineers deal with this reality through tolerances -- specified ranges of acceptable variation. A tolerance is, by definition, a statement of satisficing. It says: this dimension must be within this range. Any value within the range is acceptable. The engineer does not specify the single "optimal" value and demand that it be achieved. The engineer specifies a range of "good enough" values and accepts whatever falls within that range.

Why? Because tolerances embody a deep engineering wisdom: the value of additional precision must exceed the cost of achieving it. A piston that fits its cylinder to within a thousandth of an inch works perfectly well. Making it fit to within a millionth of an inch would not make the engine noticeably better but would multiply the manufacturing cost by a factor of a hundred.

This principle extends beyond manufacturing. Software engineers practice it when they define "acceptable" response times (the page must load in under two seconds -- not as fast as theoretically possible, but fast enough that users do not notice). Architects practice it when they specify building materials (the beam must support twice the expected load -- not infinity times the expected load, but enough to provide a safety margin). Pharmaceutical companies practice it when they define therapeutic windows (the drug must maintain a blood concentration between 10 and 20 micrograms per milliliter -- not exactly 15, but somewhere in the effective range).

In every case, the tolerance is an explicit acknowledgment that perfection is unachievable and that the cost of pursuing it exceeds its value. Tolerances are satisficing encoded in technical specifications.

There is a deeper lesson here that connects to the structure of optimization landscapes we explored in Chapter 7. In gradient descent, getting close to the minimum is easy. Getting to the exact minimum is hard, and getting arbitrarily close requires increasingly many iterations with diminishing returns. The loss landscape near the minimum is flat -- many points are nearly as good as the best one. This means that the difference between "close to optimal" and "exactly optimal" is negligible in terms of performance but enormous in terms of cost. Tolerances exploit this flatness. They say: anywhere in this flat region near the optimum is fine. Do not waste resources searching for the exact bottom of a valley when the entire valley floor is effectively equivalent.

12.7 The Paradox of Choice

In 2000, the psychologists Sheena Iyengar and Mark Lepper published a study that would become one of the most cited experiments in behavioral economics. They set up a display table in a gourmet grocery store offering samples of jam. On some days, the table displayed twenty-four varieties of jam. On other days, it displayed six. Both displays attracted shoppers -- the large display attracted slightly more initial attention. But the purchasing behavior was dramatically different. Of the shoppers who stopped at the small display, 30 percent bought a jar. Of those who stopped at the large display, only 3 percent bought a jar.

More choice led to less action. This is the paradox of choice, a term popularized by the psychologist Barry Schwartz in his 2004 book of the same name. Schwartz drew on a growing body of research showing that expanding options beyond a certain point does not increase satisfaction -- it decreases it. More choice leads to more difficulty choosing, more anxiety about choosing wrong, more regret after choosing, and less satisfaction with whatever is chosen.

Schwartz distinguished between two types of decision-makers: maximizers and satisficers. Maximizers seek the best possible option. They compare exhaustively, research relentlessly, and evaluate every alternative before committing. Satisficers seek an option that is good enough. They define their criteria, search until they find something that meets those criteria, and stop.

The research consistently shows that satisficers are happier. They experience less regret, less anxiety, and greater satisfaction with their choices. Maximizers, despite putting in more effort and often making objectively "better" choices (by some external metric), are less satisfied because they are haunted by the options they did not choose. The maximizer who buys the best-reviewed washing machine is less happy than the satisficer who buys a perfectly adequate one, because the maximizer is aware of -- and tormented by -- the seventeen other washing machines she considered and rejected.

This finding has a direct connection to the explore/exploit tradeoff from Chapter 8. Maximizers are, in essence, pure explorers -- they keep searching, keep comparing, keep exploring the option space long past the point of diminishing returns. They exploit too little, committing to a choice too late or never fully committing at all. Satisficers exploit early -- they find something good enough and exploit it, freeing their time and cognitive resources for other decisions.

The paradox of choice also connects to the information-theoretic ideas from Chapter 6. Each additional option adds information to the decision environment. But when the number of options exceeds the decision-maker's capacity to process them, the additional information becomes noise, not signal. Twenty-four jams do not provide four times as much useful information as six jams. They provide roughly the same amount of useful information (there are jams in various flavors at various price points) buried in four times as much noise (trivial differences among nearly identical options). The satisficer, by imposing a threshold and stopping early, implicitly filters the noise. The maximizer, by attempting to process all available information, drowns in it.

Spaced Review -- Chapter 8 (Explore/Exploit): Recall that the optimal explore/exploit balance shifts over time. Early on, when you know little, exploration is valuable. Later, when you have learned enough, exploitation is more efficient. How does this map onto the maximizer/satisficer distinction? In what sense is a maximizer someone who never makes the transition from exploration to exploitation? What does the research on the paradox of choice suggest about the costs of over-exploration?

🔄 Check Your Understanding

1. Why does the relationship between precision and cost follow an exponential rather than a linear pattern in engineering?
2. In the jam study, why did more options lead to fewer purchases? Frame your answer in terms of satisficing.
3. How does the maximizer/satisficer distinction relate to the explore/exploit tradeoff from Chapter 8?

12.8 Fast-and-Frugal Heuristics: Less Can Be More

Herbert Simon coined the term "satisficing" and argued that heuristics are the tools of bounded rationality. But it was Gerd Gigerenzer, the German psychologist whose work on natural frequencies we encountered in Chapter 10 (Case Study 1), who showed that heuristics are not merely acceptable substitutes for optimization -- they can actually outperform it.

Gigerenzer and his research group at the Max Planck Institute for Human Development developed a research program around what they call fast-and-frugal heuristics -- simple decision rules that use minimal information, minimal computation, and minimal time, yet often produce better outcomes than complex optimization methods. The key concept in Gigerenzer's framework is ecological rationality: a heuristic is not good or bad in the abstract. It is good or bad relative to the environment in which it is used. A simple rule that exploits the structure of its environment can outperform a sophisticated model that ignores that structure.

The canonical example is the recognition heuristic. When asked "Which city is larger, Detroit or Milwaukee?", American students generally answer correctly (Detroit). When asked "Which city is larger, Detroit or Milwaukee?", German students also generally answer correctly -- but by a completely different mechanism. American students know both cities and must weigh various pieces of information. German students, who typically have heard of Detroit (from news about the auto industry, Motown, or urban decline) but not Milwaukee, use a simpler rule: "If I have heard of one city but not the other, the one I have heard of is probably larger." This heuristic -- pick the recognized option -- works because city size is correlated with media coverage, which is correlated with recognition. The heuristic exploits the information structure of the environment.

Remarkably, German students using the recognition heuristic answer correctly more often than American students using their fuller knowledge. This is the less-is-more effect: ignorance, combined with a smart heuristic, can outperform knowledge. The reason is that the additional information available to Americans introduces noise alongside signal. Some of that extra information is misleading, conflicting, or irrelevant. The recognition heuristic, by ignoring all of this and relying on a single cue, avoids the noise.

Gigerenzer's group catalogued an entire adaptive toolbox of fast-and-frugal heuristics:

• Take-the-best: When comparing options on multiple criteria, examine criteria one at a time in order of importance. As soon as one criterion discriminates between the options, choose the option that wins on that criterion and ignore all remaining criteria. This heuristic ignores most available information -- and in many environments, it outperforms regression models that use all available information.

• Tallying: Count the number of positive cues for each option and choose the option with the most. Do not weight the cues. Do not compute interactions. Just count. This is even simpler than take-the-best and sometimes works even better.

• 1/N rule: When allocating resources (say, investment across N funds), allocate equally to all options. Do not try to optimize the portfolio. This rule -- which uses zero information about the individual options -- has been shown to match or outperform Markowitz mean-variance optimization (the Nobel Prize-winning method for portfolio allocation) in many real-world settings, because the estimation errors in the optimal solution are larger than the errors introduced by equal allocation.

How can less information and simpler rules lead to better outcomes? The answer lies in a concept that will become the focus of Chapter 14: overfitting. A complex model with many parameters can fit the training data perfectly but generalize poorly to new data. It captures not only the underlying pattern but also the noise specific to the particular dataset. A simpler model, by having fewer parameters to fit, is forced to capture only the strongest patterns -- the ones most likely to hold up in new situations. Fast-and-frugal heuristics, by design, have very few parameters. They cannot overfit. This gives them an advantage in environments that are noisy, unstable, or characterized by small samples -- which is to say, in most real-world environments.

Connection to Chapter 10 (Bayesian Reasoning): Gigerenzer's fast-and-frugal heuristics might seem to contradict the Bayesian framework we explored in Chapter 10. Bayes' theorem says you should use all available evidence to update your beliefs. Fast-and-frugal heuristics say you should often ignore most of the evidence. But the contradiction is only apparent. In environments with reliable data and large samples, full Bayesian updating is optimal. In environments with noisy data, small samples, and uncertain cue validities, the estimation errors in the Bayesian calculation can swamp its theoretical advantage. The fast-and-frugal heuristic, by using less information, makes smaller errors. The two approaches are not competing theories of rationality. They are tools for different environments -- and knowing which environment you are in is the deeper skill.

12.9 Recognition-Primed Decision Making: How Experts Satisfice

If heuristics can outperform optimization for simple decisions like which city is larger, what about high-stakes, time-critical decisions where getting it right really matters? The psychologist Gary Klein spent decades studying exactly this question, embedded with firefighters, military commanders, intensive care nurses, and other professionals who make life-and-death decisions under extreme time pressure.

Klein expected to find that experts make decisions by comparing options -- weighing pros and cons, evaluating alternatives, choosing the best one. He found the opposite. Experts almost never compare options. They use a process Klein called recognition-primed decision making (RPD).

Here is how it works. An experienced fireground commander arrives at a burning building. He does not sit down and list his options (enter through the front, enter through the side, set up a defensive perimeter, call for additional units). He does not weigh each option against the others. Instead, he looks at the situation -- the color of the smoke, the behavior of the flames, the construction of the building, the wind direction -- and recognizes it as an instance of a pattern he has seen before. The pattern activates a course of action that has worked in similar situations. The commander mentally simulates that action: "If I send the crew in through the front, will the floor hold? Will the stairway be passable?" If the simulation reveals a fatal flaw, he modifies the plan or generates another. But he does not compare alternatives. He satisfices: he finds the first course of action that seems workable and goes with it.

Klein's research found that experienced decision-makers use this process roughly 80 percent of the time. They generate a single option, test it mentally, and either adopt it or modify it. They almost never generate multiple options and compare them. This is satisficing in its purest expert form: the expert's extensive experience has built a library of patterns so rich that the first option generated is usually good enough. The cost of generating and comparing additional options -- in time, cognitive load, and delayed action -- exceeds the marginal benefit.

This finding has radical implications for how we think about expertise. The classical view holds that experts are better optimizers -- they consider more options, weigh them more carefully, and choose more precisely. Klein's research suggests the opposite: experts are better satisficers. Their expertise consists not in superior comparison but in superior pattern recognition -- the ability to see a situation, match it to an appropriate response, and act without deliberation. The expert's speed and accuracy come from recognizing what the situation is, not from computing what the optimal action would be.

This connects beautifully to the Bayesian framework from Chapter 10. The expert's pattern library functions as a rich prior. When a new situation arises, the expert does not reason from scratch -- she matches the situation to her extensive prior experience, which effectively narrows the hypothesis space to a single plausible response. This is Bayesian updating with an extremely informative prior: when your prior knowledge is strong enough, you do not need to compare alternatives. The first hypothesis your prior generates is overwhelmingly likely to be adequate.

Spaced Review -- Chapter 10 (Bayesian Reasoning): In Chapter 10, we discussed how priors are not bias but accumulated knowledge that makes reasoning more efficient. How does Klein's research on recognition-primed decision making illustrate this principle? What serves as the "prior" in a firefighter's rapid assessment of a burning building? Why is this prior an advantage rather than a bias?

12.10 When Optimization Fails: Three Warnings

We have seen that satisficing is often better than optimization because optimization is computationally intractable, because the world is too uncertain for precise optimization to be reliable, and because simple strategies can outperform complex ones in noisy environments. But there are three specific failure modes of optimization that deserve particular attention, because they recur across every domain this book examines.

Warning 1: Overfitting

When you optimize too precisely for the data you have, you fit the noise as well as the signal. The result is a model, strategy, or plan that performs brilliantly on past data and terribly on future data. We touched on this in Section 12.8 when discussing why fast-and-frugal heuristics outperform complex models. Chapter 14 will explore overfitting in depth, but the connection to satisficing is worth stating now: satisficing is a natural defense against overfitting. By accepting a "good enough" solution rather than the "best" solution for the current data, you are implicitly building in a margin that absorbs future variation. The satisficed solution may not be the tightest fit to today's data, but it is more likely to still work tomorrow.

Think of it this way. An engineer who specifies a tolerance of plus or minus one millimeter is satisficing -- accepting any part within that range. This means the part will still function even if conditions change slightly (thermal expansion, wear, variation in mating parts). An engineer who specifies the exact theoretical optimum -- 10.000 centimeters with zero tolerance -- has created a part that works perfectly under exactly one set of conditions and fails under all others. The tolerance, the satisficed range, is what gives the system robustness.

Warning 2: Goodhart's Law

"When a measure becomes a target, it ceases to be a good measure." This principle, attributed to the British economist Charles Goodhart and sometimes stated as Campbell's Law, describes what happens when you optimize too aggressively for a specific measurable metric.

A hospital wants to reduce patient wait times. It sets a target: no patient should wait more than four hours in the emergency department. This target is measurable, clear, and seems like a reasonable optimization goal. But when the hospital optimizes for it, strange things happen. Patients are admitted to inpatient wards prematurely to get them off the emergency department clock. Triage categories are manipulated. Ambulances are told to circle the block. The metric improves. The patient experience does not.

This is Goodhart's Law in action, and it is a failure of optimization. The metric (wait time) was a proxy for the real goal (patient welfare). When you optimize the proxy, you inevitably find ways to improve the proxy that do not improve -- and may actively harm -- the underlying goal. The proxy and the goal diverge under optimization pressure.

Satisficing avoids this trap, or at least mitigates it. If the hospital satisfices -- "wait times should be reasonable, roughly under four hours for most patients" -- there is no incentive to game the specific metric because there is no specific metric to game. The satisficing criterion is vague enough to resist Goodhart's Law precisely because it is too imprecise to optimize. Chapter 15 will explore Goodhart's Law in detail; for now, note that it is one of the strongest arguments for satisficing over optimization in any complex system.

Warning 3: Computational Intractability

We discussed this in Section 12.4, but it bears emphasis. Many optimization problems are not just hard in practice -- they are provably hard in theory. No algorithm can solve them efficiently as the problem scales up. The traveling salesman problem, for instance, has no known polynomial-time solution. For a salesman visiting 20 cities, the optimal route can be found by brute force. For 200 cities, the brute-force approach would take longer than the age of the universe. For 2,000 cities, the numbers become so large that they lose all physical meaning.

In practice, the traveling salesman problem is solved by heuristics -- nearest-neighbor algorithms, simulated annealing, genetic algorithms -- that find good-enough routes without guaranteeing optimality. These heuristics are satisficing strategies. They accept a route that is within a few percent of the theoretical optimum, at a computational cost that is within a few seconds rather than a few billion years. The few percent of quality sacrificed is the price of actually having an answer.

🔄 Check Your Understanding

1. How is satisficing a natural defense against overfitting?
2. Explain Goodhart's Law in your own words and give an example from education, business, or government.
3. Why does computational intractability make satisficing not just practical but theoretically necessary for many optimization problems?

12.11 The 80/20 Connection: Power Laws and Satisficing

In Chapter 4, we explored power laws -- distributions where a small number of items account for a disproportionate share of the total. The Pareto principle, or 80/20 rule, is the most familiar version: roughly 80 percent of effects come from roughly 20 percent of causes. Eighty percent of sales come from 20 percent of products. Eighty percent of bugs come from 20 percent of the code. Eighty percent of a language's usage consists of roughly 20 percent of its vocabulary.

The 80/20 rule has a direct and profound connection to satisficing. If 80 percent of the value comes from 20 percent of the effort, then the first 20 percent of effort is twenty times as productive as the remaining 80 percent. The satisficer who does the first 20 percent and stops captures 80 percent of the value at 20 percent of the cost. The optimizer who insists on the full 100 percent invests five times as much effort for a 25 percent improvement in outcome.

This arithmetic explains why satisficing is not just psychologically comfortable but economically rational. In any domain governed by diminishing returns -- which is nearly every domain -- the marginal value of additional effort decreases while the marginal cost stays constant or increases. At some point, the cost of improvement exceeds its value. That point is the satisficing threshold.

The connection to power laws goes deeper. In power-law distributed environments, the variation among the "top" options is enormous, while the variation among the "good" options is small. The difference between the best cereal and the second-best cereal (by whatever metric you choose) may be noticeable. The difference between the twelfth-best and the twentieth-best is negligible. This means that satisficing -- picking any option from the broad pool of "good enough" choices -- costs you very little in quality. The returns to further search are concentrated at the very top of the distribution, and the probability of landing there through additional search is low (because, by the nature of power laws, there are very few options at the extreme).

This is why Schwartz's research shows that maximizers and satisficers make choices of similar objective quality. The distribution of options is such that many options are nearly equally good, and the marginal improvement from exhaustive search is tiny. The maximizer's additional effort buys almost nothing in quality -- but costs a great deal in time, stress, and regret.

Pattern Library Checkpoint: We have now seen satisficing -- the strategy of finding a "good enough" option rather than the optimal one -- in evolution (the backwards retina, the recurrent laryngeal nerve), chess engines (heuristic evaluation, pruning), military strategy (Auftragstaktik, mission-type tactics), engineering (tolerances, safety margins), consumer psychology (the paradox of choice, maximizers vs. satisficers), expert decision-making (recognition-primed decisions), cognitive science (fast-and-frugal heuristics), and now economics (the 80/20 rule). Add this to your pattern library: satisficing is not a failure mode but a design principle, and it emerges wherever decision-makers face environments that are too complex, too uncertain, or too costly to optimize.

12.12 The Threshold Concept: Bounded Rationality Is Not Irrationality

We have arrived at the deepest idea in this chapter, the one that most students initially resist and that marks a genuine shift in understanding once it is grasped.

The standard view of rationality -- the view inherited from economics, decision theory, and much of philosophy -- equates rationality with optimization. A rational agent is one who computes the best possible action given all available information and acts accordingly. Deviations from this standard -- choosing a suboptimal option, using a heuristic instead of a full calculation, stopping search before all alternatives have been considered -- are classified as failures of rationality, cognitive biases, or "irrational" behavior.

Simon's insight was that this view is exactly backwards.

If optimization is impossible -- because the information is incomplete, the computation is intractable, or the time is insufficient -- then defining rationality as optimization is like defining health as immortality. It sets an unachievable standard and then classifies all actual behavior as pathological. Under this definition, every decision ever made by every human being in history is "irrational," because no human being has ever had the information, computational capacity, or time to fully optimize any non-trivial decision.

Bounded rationality redefines rationality in terms of what is actually achievable. A boundedly rational agent is one who uses strategies that are well-adapted to the actual constraints of the decision environment -- limited information, limited computation, limited time. Satisficing is not a failure to optimize. It is the rational response to an environment in which optimization is impossible. Using a heuristic is not a cognitive shortcut born of laziness. It is an ecologically rational strategy -- one that exploits the structure of the environment to achieve good outcomes with minimal resources.

This reframing has far-reaching consequences. It means that many behaviors classified as "cognitive biases" by the heuristics-and-biases tradition (associated with Daniel Kahneman and Amos Tversky, whose work we will encounter in Chapter 22) may not be biases at all. They may be ecologically rational heuristics -- strategies that perform well in the environments they evolved to handle, even if they fail in the artificial environments of laboratory experiments and logic puzzles. The recognition heuristic, for instance, looks like a "bias" if you think rational agents should use all available information. It looks like ecological rationality if you recognize that in many natural environments, recognition is a reliable cue and additional information is noisy.

This does not mean that all heuristics are good or that humans never make mistakes. It means that the standard against which we judge rationality matters. If the standard is classical optimization, then everything looks like a failure. If the standard is adaptive fit to the decision environment, then many "biases" look like features, and the question becomes not "why are humans irrational?" but "in which environments do human heuristics work well, and in which do they break down?"

Gigerenzer calls this ecological rationality: the study of the match between decision strategies and decision environments. A heuristic is ecologically rational when its assumptions match the structure of the environment. The recognition heuristic is ecologically rational in environments where recognition is correlated with the criterion of interest (city size, product quality). It is ecologically irrational in environments where recognition is manipulated (advertising, propaganda) or uncorrelated with quality (random domains).

The deepest lesson is this: there is no universally optimal decision strategy. There is no algorithm that is best in all environments. What exists is an adaptive toolbox -- a repertoire of strategies, some simple, some complex, some fast, some slow -- and the skill of selecting the right strategy for the situation at hand. The boundedly rational agent is not the one who uses the most sophisticated strategy. It is the one who matches strategy to environment -- who knows when to satisfice and when to invest in more careful analysis, when to use a heuristic and when to reach for a more complex model.

This is, in the end, what cross-domain pattern recognition is about. The satisficing pattern -- the recognition that "good enough" is a legitimate and often superior standard -- appears in every domain we have examined. The skill is not in learning the pattern once but in recognizing it everywhere, seeing that the engineer's tolerance and the firefighter's rapid recognition and the bacterium's run-and-tumble and the chess engine's pruning algorithm are all instances of the same deep structure: the adaptive response of intelligent systems to the impossibility of optimization.

🔄 Check Your Understanding

1. In your own words, explain why Simon argued that bounded rationality is not irrationality.
2. What is ecological rationality? Give an example of a heuristic that is ecologically rational in one environment but ecologically irrational in another.
3. Why is there no universally optimal decision strategy? What does this imply about the skill of good decision-making?

12.13 The Satisficing Landscape: A Synthesis

Let us step back and view the full landscape we have traversed.

We began in the cereal aisle, watching a real human make a real decision by scanning, satisficing, and moving on. We then traced the same pattern across domains:

In every case, the satisficing strategy sacrifices theoretical perfection for practical robustness, speed, simplicity, and adaptability. In every case, the sacrificed perfection was unachievable anyway. And in several cases -- fast-and-frugal heuristics, the 1/N rule, the recognition heuristic -- the satisficing strategy actually outperforms the theoretically optimal approach, because the optimal approach overfits, overcomplicates, or overestimates the reliability of available information.

The meta-pattern is this: in complex, uncertain, resource-constrained environments, simplicity is not a compromise. It is a competitive advantage. The system that finds a good-enough solution fast will outperform the system that searches for the perfect solution slowly, because the perfect solution does not exist, the search is costly, and the environment changes while you are searching.

This is the view from everywhere. The evolutionary biologist, the chess programmer, the military strategist, the cognitive psychologist, the engineer, and the behavioral economist are all looking at the same elephant. They have given it different names -- fitness, evaluation function, mission-type tactics, bounded rationality, tolerance, satisficing. But the underlying structure is identical: the adaptive superiority of "good enough" in a world where "optimal" is either unattainable, unreliable, or too expensive to compute.

Forward Reference: In Chapter 14 (Overfitting), we will explore in depth why optimizing too precisely is dangerous -- how fitting too closely to available data causes systems to fail on new data. Overfitting is, in a precise sense, the pathology that satisficing prevents. In Chapter 15 (Goodhart's Law), we will examine how optimizing the wrong metric leads to perverse outcomes -- how measuring and maximizing a proxy for what you actually want invariably distorts the proxy until it no longer measures what you care about. Both chapters will draw on the satisficing framework developed here as the antidote to these optimization pathologies.

12.14 Living With Good Enough

There is a personal dimension to satisficing that goes beyond theory.

We live in a culture that celebrates optimization -- the best workout, the best diet, the best productivity system, the best school for your children, the best investment strategy, the best life. Social media amplifies this by making every choice visible and every alternative accessible. You can see what everyone else chose, and it always looks better. The implicit message is relentless: you should be optimizing. You should be choosing the best. Anything less is settling.

Simon's research suggests that this cultural pressure is not just impractical but actively harmful. The pursuit of the best is a recipe for paralysis (too many options to evaluate), regret (every chosen option forecloses others), and dissatisfaction (the optimal choice, even if found, feels inadequate because you are aware of how close the alternatives were). The satisficer who chooses "good enough" and moves on is freed from all three.

This does not mean you should not care about quality. It means you should care about quality thresholds rather than quality rankings. "Is this cereal good enough? Does it taste acceptable, meet my nutritional requirements, and cost a reasonable amount?" is a question you can answer. "Is this the single best cereal among three hundred options?" is a question that will destroy your afternoon and leave you holding a box of granola you are no longer sure you want.

The skill is in setting the right threshold. Too low, and you settle for genuinely bad options. Too high, and you are back to maximizing under a different name. The art of satisficing is calibrating your aspiration level to match the stakes of the decision. For breakfast cereal, the threshold can be low. For a career, a life partner, or a surgical procedure, the threshold should be higher -- but it should still be a threshold, not a demand for global optimization.

Herbert Simon, who spent his career studying the limits of rationality, seems to have practiced what he preached. His colleagues described him as a man who made decisions quickly, committed to them fully, and rarely looked back. He did not agonize over the optimal research topic -- he worked on whatever seemed most interesting and important. He did not search for the perfect word -- he wrote clearly and moved on. He satisficed, and in doing so, he produced one of the most remarkable bodies of work in the history of the social sciences.

Perhaps the deepest irony of satisficing is this: the people who accept "good enough" often end up with lives that are, by any reasonable measure, better than those of the people who insist on the best. Not because they are lucky. Because they are free -- free from the impossible task of optimization, free from the paralysis of infinite comparison, free from the regret of paths not taken. They have made peace with the fundamental human condition: we are finite creatures in an infinite world, and the art of living well is the art of choosing wisely within our limits.

That is bounded rationality. It is not irrationality. It is wisdom.

Chapter Summary

Herbert Simon's concept of satisficing -- searching for an option that is "good enough" rather than computing the theoretically optimal one -- is not a failure of rationality but an adaptive response to the real-world constraints of limited time, information, and computational capacity. This chapter traced the satisficing pattern across evolution (which satisfices by retaining any variation that clears the survival threshold), chess engines (which use heuristics rather than exhaustive search), military strategy (which favors adaptive planning over rigid optimization), engineering (which specifies tolerances rather than demanding perfection), consumer psychology (where satisficers are happier than maximizers), expert decision-making (where pattern recognition replaces option comparison), and cognitive science (where fast-and-frugal heuristics often outperform complex optimization).

The chapter's threshold concept -- Bounded Rationality Is Not Irrationality -- reframes rationality as adaptive fit to the decision environment rather than conformity to an impossible standard of optimization. In complex, uncertain, resource-constrained environments, simplicity is not a compromise but a competitive advantage, and the skill of good decision-making lies not in computing the best answer but in matching the decision strategy to the decision environment.