David Thorstad (Global Priorities Institute & Kellogg College, University of Oxford), "The accuracy-coherence tradeoff in cognition"

Forthcoming, British Journal for the Philosophy of Science

By David Thorstad

Humans are bounded agents. We have limited cognitive abilities, and in exercising our abilities we incur costs.

The lives of bounded agents are fraught with tradeoffs. For example, often (not always!) we face accuracy-effort and speed-accuracy tradeoffs: more accurate belief-forming rules tend to require more effort and time to apply. These tradeoffs are standardly thought to have implications for how bounded agents should cognize. For example, in many areas we may use belief-forming rules whose reliability is satisfactory, but not best-possible in order to preserve time and resources for other tasks.

In this paper, I argue that there is a comparably systematic accuracy-coherence tradeoff for bounded agents. In many circumstances, the cognitive processes which produce the most accurate judgments come apart from the processes which produce the most coherent judgments. As a result, bounded agents are often forced to choose whether to privilege accuracy or coherence in belief-formation.

I use a case study of lexicographic and semilexicographic methods in judgment and decisionmaking to illustrate how the accuracy-coherence tradeoff can arise. Then I identify and study two of the many factors driving the accuracy-coherence tradeoff: cognitive complexity and strategic variety.

I conclude by discussing normative implications of the accuracy-coherence tradeoff. The accuracy-coherence tradeoff reveals that the pursuit of coherence can come at the expense of predictive accuracy. This raises the question of how bounded agents should balance accuracy and coherence in cognition.

Many philosophers defend approximate coherentist theories of bounded rationality, on which agents should strive to be as coherent as possible given their limitations. Approximate coherentism takes an extremal approach to the accuracy-coherence tradeoff, on which no amount of increased accuracy can ever compensate for decreased coherence. But insofar as rational agents also strive to be accurate, we may want to develop more moderate theories of bounded rationality which allow us to weigh competing cognitive desiderata, such as accuracy and coherence, against one another.

This does not require us to deny that coherence has value. But it may suggest that bounded rationality cannot be exhausted by coherence, because competing goals such as accuracy have value too.

You can read and listen to a short version of the paper at BJPS Short Reads, or read the full paper here.