The result is: the performance of “Take the Best” is equivalent to a linear regression model with noncompensatory cues. The more condition (1) is violated, the more the performance of the heuristic may deteriorate.

Chapters 7 and 8 focus more on the psychology of decision making rather than on the performance of particular heuristics, namely trying to understand when people use heuristics, and a comparison of heuristics to another model of decision making, the Bayesian network. I took two messages away from these chapters, namely, (1) the more people are pressed for time, the more they use heuristics; and (2) the performance of heuristics does not deteriorate as much as the Bayesian network when generalized.

The idea that people use heuristics when they are pressed for time seems to undermine the book's claim that people do not use complex algorithms to solve problems. Chapter 7 suggests that the more time people have to think, the more they try to understand correlations, integrate across cues, and use other complicated decision algorithms. Indeed, the authors refer to a study demonstrating that “participants who have to account for their decisions more frequently select strategies that require considerable effort but offer a higher probability of a correct decision.” Near the end of Chapter 7, the authors state that “Even if people use a large amount of information and do integrate it, they seem to use simple cognitive operations.” This chapter failed to convince me that there is a meaningful separation between more complex cognitive processes and heuristics.

Chapter 8 highlights a strong point of heuristics: they are less subject to overfitting than more complex algorithms. A Bayesian network falls prey to this when it is trained on a subpopulation in which correlations are stronger than in the population from which it was drawn. The way the text is written suggests that this result should be surprising. As they put it, “Simple one-reason decision making, as employed by ‘Take the Best,’ is almost as accurate as the computationally expensive Bayesian network.” Note that the Bayesian network does do better in most cases than TTB, but not by much. I was unconvinced that this was not the result of a mismatch between the complexity of the prediction and the complexity (or cost) of algorithm used to make it.

Chapters 9–14 focus on more complex decision processes than binary choice, namely memory, estimation, and categorization, inferring intention from motion, mate selection, and parental investment. Chapter 9 explores how “hindsight bias” (the tendency to believe falsely, after the fact, that one would have predicted the outcome of an event) may be the by-product of a heuristic for memory updating and inference, rather than an error of human information processing. Chapter 10 develops a heuristic to estimate the populations of cities that exploits the nature of the j-shaped underlying distribution. Chapter 11 compares a heuristic for Categorization by Elimination with more complex algorithms including a neural network. Each of these chapters provides a nice connection to the existing literature, allowing the tasks to be put in context.

Chapter 12 compares the performance of the “Categorization by Elimination” (CBE) heuristic to more complex algorithms, including a neural network to categorize intent (pursuit, evasion, fighting, courting, being courted, and playing) from motion cues (relative distance, relative angle, relative heading, absolute velocity, relative velocity, absolute vorticity, and relative vorticity) of two bugs simulated on a computer. What struck me here was that the difference in performance between the neural network and CBE was more marked than in other cases. The discussion focuses on how CBE does almost as well as two other linear integrative strategies: Dawe's rule which simply adds up the cue values, and Franklin's rule which is a weighted sum with the weights based on the data (training set), with less information. I would have found a discussion on why the neural net did much better in this case very interesting, i.e., what is it about the environment in this example that improved the performance of the neural network?

Chapters 13 and 14 apply heuristics to topics prevalent in the biological literature: mate choice and parental investment. In Chapter 13, a variant of TTB is pitted against the optimal rule for a simple model of mate selection. The optimal rule guarantees the highest chance of picking the best candidate. If you are interested in picking not the best, but within the top 25% of candidates, the heuristic performed well. As with previous examples, for coarse-grained predictions, heuristics perform well. In Chapter 14, six simple heuristics for feeding chicks were discussed. The best of these simple heuristics was then compared to three constrained optimization models. According to the authors, the simple heuristic outperformed all three optimization models. The importance of this result was diminished by the fact that the optimization models to which the heuristic was compared were somewhat ad hoc.

The final two chapters attempt to pull together all of the lines of thought developed in the book. Chapter 15, titled “Demons Versus Heuristics in Artificial Intelligence, Behavioral Ecology, and Economics” explores the role heuristics might play in understanding behavior in these fields. What caught my attention most in this section was the discussion of rationality in animals. The authors discuss optimization techniques in foraging theory and the deviation of actual behavior from “optimal.” The author states that “If behavioral ecologists have always been aware of constraints [which cause deviations from actual behavior] they have not always taken them into account in trying to explain behavior.” I question the accuracy of this statement in general. In my experience with stochastic dynamic programming models in behavioral ecology, a careful discussion of constraints and their influence on the explanation of behaviour is an essential part of the modeling process.

The authors go on to suggest that the inclusion of constraints into optimization models falls short of “true” models of bounded rationality such as “satisficing.” In the latter, organisms are not assumed to be optimizing, but rather, attempting to achieve a minimum threshold. Any improvement beyond this is comparatively unimportant. The importance of the separation between heuristics and constrained optimization is overstated. For one, it is straightforward to alter the objective function to incorporate thresholds (i.e., the Stone-Geary utility function in economics), and thus capture satisficing-like behavior in an optimizing framework. Next, the heuristic itself might well be the optimal solution for the given problem. I do not see the value or importance of making such a strong separation here. Indeed, the authors later note that “One promising explanation is that aspiration levels and rules of thumb themselves are the result of evolutionary selection.” As with my earlier comment regarding behavioral ecology, heuristics may be (approximate) solutions to optimization problems given physiological constraints and historical contingencies.

In summary, although the book provides some interesting discussions of decision processes, it is too long. Too many examples are explored that share the same basic structure, i.e., that in particular environments where the aim is a coarse-grain estimate of some criterion, heuristics can be effective. This seems to me to be a simple example of the trade-off between the value of information and the cost of obtaining it. No examples were presented in which heuristics performed very poorly. Those that are presented are not related to well-developed, standard notation from probability and statistics, i.e., the examples are about predicting random variables, given a set of independent variables. The increased predictive power of additional independent variables in a model will depend on the correlation between those variables. This is all well known. A more direct approach of investigating under what circumstances heuristics perform well in the context of well-developed theory would have been more enlightening than the approach of the book: the presentation of several examples in which heuristics perform well as surprising and counterintuitive. Finally, the authors make too much of the distinction between heuristics and constrained optimization. I was not convinced that heuristics provide a better model for human (or animal) behavior than do other more complex models such as Bayesian networks or constrained optimization. I was left to believe that people may apply a range of decision processes (including complex ones that integrate cues) to different problems in different environments. Whether we are conscious that we used a more complex decision rule, perhaps one hard-wired by natural selection, is another story. I expect that few Conservation Ecology readers would find this book very useful in their research. Those interested in behavioral ecology, human sociobiology, or psychology might find the perspective presented in the book of some interest.

Gigerenzer, G., P. M. Todd, and the ABC [Adaptive Behavior and Cognition] Research Group. 2000. Simple Heuristics that Make Us Smart. Oxford University Press, Oxford, UK. 416 pp., paperback, US$19.95. ISBN 0195143817.

Responses to this article are invited. If accepted for publication, your response will be hyperlinked to the article. To submit a comment, follow this link. To read comments already accepted, follow this link.

I would like to thank an anonymous referee for comments that significantly improved this book review.

Gigerenzer, G., P. M. Todd, and the ABC Research Group. 2000. Simple heuristics that make us smart. Oxford University Press, Oxford, UK.