Strategic Reliablism

post by badger · 2011-05-25T17:33:03.814Z · LW · GW · Legacy · 0 comments

Contents

  Ch 1. Laying Our Cards on the Table
  The Amazing Success of Statistical Prediction Rules
  Extracting Epistemic Lessons from Ameliorative Psychology
None
No comments

 

Epistemology and the Psychology of Human Judgment by Michael Bishop and J.D. Trout

---

Ch 1. Laying Our Cards on the Table

Epistemology as a discipline needs to start offering practical advice for living. Defective epistemologies can compromise one's ability to act in all areas, but there is little social condemnation of weak reasoning. Prescriptive epistemology might be called "critical thinking", but this field is divorced from contemporary epistemology. This book is driven by a vision of what epistemology could be, but is, of course, only a modest first step in that direction.

Standard Analytic Epistemology (SAE) is primarily concerned with an account of knowledge and epistemic justification. This program assumes any account of justification must not radically alter our existing judgments, even if this commitment to stasis is not explicitly stated. SAE tends to unproductively proceed via counterexamples. SAE might provide some useful advice, but it's goals and methods are beyond repair.

In the authors' view, epistemology is a branch of philosophy of science and should start from Ameliorative Psychology, which encompasses parts of cognitive science, the heuristics and biases program, statistics, and artificial intelligence. This field makes recommendations about how to reason based on empirical findings. For example, statistical prediction rules (SPRs) were found to be at least as reliable as human experts and frequently more so. As another example, Bayesian reasoning is more successful when information is presented in frequencies rather than probabilities. Rather than being concerned with an account of knowledge or warrant, the authors' approach is to provide an account of reasoning excellence.

A healthy epistemological tradition will have theoretical, practical, and social components. Theory and practice should mutually inform one another. Communication of practical results to the wider public is an important aspect of the social component. Since the authors argue epistemology is a science, but nevertheless normative, one might worry the approach is circular. This concern assumes the normative must come all at once, though. Instead, one can rely on the Aristotelian Principle as an empirical hook to accumulate justification. The Aristotelian Principle says that in the long run, good reasoning tends to lead to better outcomes than poor reasoning. Why accept the Principle? Without it, epistemology wouldn't be useful. If bad reasoning leads to better outcomes and if there are many types of bad reasoning, how could we figure out which bad reasoning to use, except by good reasoning? We have reason to think useful epistemology is possible since we live in a stable enough environment where quality has a chance to make a difference.

The Amazing Success of Statistical Prediction Rules

Judgments are an essential part of life. Choices of whether to release a prisoner on parole, admit someone to medical school, or offer a loan are too important to be "close enough". Only the best reasoning strategies available to us are satisfactory. Statistical prediction rules are robustly successful in these and many other high-stakes areas. In 136 studies comparing proper linear models to expert judgment, 64 clearly favored the SPR, 64 showed statistically equivalent accuracy, and 8 favored the expert. ¹ SPRs built explicitly to mimic experts' judgments are more reliable than the expert, suggesting at least some errors are due to inconsistency and making exceptions to one's own rules.

Improper linear models with unit or even random weights on standardized variables do surprisingly well. Qualitative human judgment can always be used as an inuput to an SPR or used to select variables and the direction of their effect. The flat maximum principle says that as long as the sign on coefficients is correct, all linear models do approximately the same. This principle applies when the problem is difficult and the inputs are reasonable predictive and redundant. Summing together inputs can be viewed as exploiting Condorcet's jury theorem. Linear models tend to work when inputs interact monotonely, which appears to be the case in most social situations.

All this is not to say SPRs are especially good, but that humans are very bad predictors. We pick up false patterns and are unable to consider even medium amounts of information at once. Resistance to SPR findings typically comes from a belief in epistemic exceptionalism. There is an impulse to tweak the conclusions of an SPR, which leads to worse results. It is surprising to find out we do so badly that random linear models can do better, yet another manifestation of overconfidence. Ironically, experts are best suited to deviate from SPRs grounded in theory because they have a better understanding of when the SPR will apply.

Extracting Epistemic Lessons from Ameliorative Psychology

Ameliorative Psychology offers a number of useful recommendations, but its normative assumptions are rarely stated explicitly. The authors identify three factors underlying the quality of a reasoning strategy: reliablity across a wide range of problems, tractability, and applicability to significant problems. Strategies need to be robustly reliable to survive changes in environments. Cheaper and easier strategies allow one to "purchase" more truths. Simple strategies like SPRs have tended to be more successful as well, possibly by avoiding overfitting, but a easy, low-quality rule is better than a high-quality one that is never used. Finally, the world is full of useless correlations, so the trick is to find important ones.

Cost-benefit relations have diminishing marginal returns. By considering possible cost-benefit curves, startup costs, and marginal expected reliability, the possible ways to improve reasoning fall into exactly four categories. Three possible ways can be seem in the following matrix:

	Same (or lower) cost	Higher Cost
Greater Benefit	(1) Adopt more reliable, cheaper strategy.	(2) Adopt more reliable, expensive strategy.
Same (or less) Benefit	(3) Adopt less reliable, but cheaper strategy.

0 comments

Comments sorted by top scores.