Belot, Gordon (2024) Mechanical Turkeys. [Preprint]

Text Turkey.pdf Download (265kB)

Abstract

Some learning strategies that work well when computational considerations are abstracted away from become severely limiting when such considerations are taken into account. We illustrate this phenomenon for agents who attempt to extrapolate patterns in binary data streams chosen from among a countable family of possibilities. If computational constraints are ignored, then two strategies that will always work are learning by enumeration (enumerate the possibilities---in order of simplicity, say---then search for the one earliest in the ordering that agrees with your data and use it to predict the next data point) and Bayesian learning. But there are many families of computable data streams that, although they can be successfully extrapolated by computable agents, cannot be handled by any computable learner by enumeration. And while there is a sense in which Bayesian learning is a fully general strategy for computable learners, the ability to mimic powerful learners comes at a price for Bayesians: they cannot, in general, become highly confident of their predictions in the limit of large data sets and they cannot, in general, use priors that incorporate all relevant background knowledge.

---

Export/Citation:

EndNote | BibTeX | Dublin Core | ASCII/Text Citation (Chicago) | HTML Citation | OpenURL

Social Networking:

Share |

---

Item Type:	Preprint
Creators:	Creators Email ORCID Belot, Gordon
Additional Information:	Forthcoming in Journal of Philosophical Logic
Keywords:	Induction; Bayesian learning; learning by enumeration; computable Bayesianism
Subjects:	Specific Sciences > Computation/Information General Issues > Confirmation/Induction Specific Sciences > Probability/Statistics
Depositing User:	Gordon Belot
Date Deposited:	29 Nov 2024 13:21
Last Modified:	29 Nov 2024 13:21
Item ID:	24294
Subjects:	Specific Sciences > Computation/Information General Issues > Confirmation/Induction Specific Sciences > Probability/Statistics
Date:	28 November 2024
URI:	https://philsci-archive.pitt.edu/id/eprint/24294

Monthly Views for the past 3 years

Monthly Downloads for the past 3 years

Plum Analytics

Actions (login required)

View Item