Abstract
The threshold view says that a person forms an outright belief P if and only if her credence for P reaches a certain threshold. Using computer simulations, I compare different versions of the threshold view to understand how they perform under time pressure in decision problems. The results illuminate the strengths and weaknesses of the various cognitive strategies in different decision contexts. A threshold view that performs well across diverse contexts is likely to be a cognitively flexible and context-dependent fusion of several of the existing theories. The results of the simulations also cast doubts on the possibility of a threshold view that is both simple enough to streamline our reasoning while also allowing us to form good action-guiding beliefs.
Similar content being viewed by others
Data availability
N/A.
Code availability
Code for the model can be found here: https://github.com/alicecwhuang/threshold_code.git.
Notes
The lottery paradox is a problem for most threshold views with \(t < 1\). Suppose there is a lottery with one winning ticket out of n tickets, where \((1-\frac{1}{n}) > t\). For each individual ticket, S’s credence that it loses suffices for an outright belief according to standard threshold views. Therefore, S outright believes that each ticket will lose. However, S also outright believes, by assumption, that at least one ticket will win. It follows that conjunctive closure is violated. In other words, if \(t < 1\), then S outright believes P and Q without believing \(P\wedge Q\).
That outright beliefs are at least partially fixed by credences is the core idea of the threshold view. Alternatives to the threshold view include, first, theories according to which graded and binary beliefs are entirely independent. This leads to prima facie strange results, for in virtue of what do the two systems both count as belief states if there is no connection between them? Secondly, graded beliefs might be fixed by binary beliefs. This approach is favored by Holton (2014) and Easwaran (2016), who think that graded beliefs are reducible to binary beliefs.
Greco (2015) later turns to support what I will call “SmallSens(1), ” or “credences-one sensitivism” and accept irrationality as an integral part of beliefs.
Being subjected to a diachronic Dutch-Book might not be a mark of irrationality in some views. See for example Christensen (1991).
Weatherson (2005) has a similar spirit of anti-idealization, though he focuses on limiting the domain where conjunctive closure should be obeyed, as opposed to limiting the set of active possibilities. He argues that we can have probabilistic coherence without logical coherence, and in particular, that only the set of salient propositions must be conjunctively closed. See also Weisberg (2020) for more on how credence is assigned on the fly, and Stanley (2007) for a normative construal of active possibilities.
See Mayseless and Kruglanski (1987) for the psychological perspective on this topic.
Babic (2019) endorses a pragmatic view similar to small domain purism, though he does not address the belief-credence connection, nor does he explicitly mention the concept of an active domain.
See Nagel (2010) for an extensive discussion about the termination of inquiries.
See Duckworth et al. (2018) 2.2.1 for a brief introduction.
The Weibull distribution is determined by two parameters—shape and scale. The shape parameter is randomized between 1 and 2, and the scale is fixed at 50. The two parameters together determine, roughly, how rapidly the patient’s probability of death increases with each time step. For instance, when the shape parameter is small, the probability of death increases faster in the beginning and slower towards the end. The smaller the scale is, the faster the patients’ probability of death increases with time. My setup of the problem here roughly follows that of Douven (2020).
The effects of correct and incorrect interventions are determined by parameters a and b. Let the patient’s baseline probability of death without intervention at a given time be p. The correct intervention brings down the probability of death to p/a, whereas a wrong intervention increases it to \(\frac{(p+1)}{b}\). The larger a is, the more effectively a correct intervention lowers the patient’s risk of death. The smaller b is, the more a wrong intervention increases the patient’s probability of death. a is randomized between 1.5 and 2.5 whereas b is fixed at 2.
The symmetric dirichlet distribution is determined by a concentration parameter. The larger the concentration parameter is, the more equally probability is distributed across the hypotheses. By contrast, when the concentration parameter is small, the agent has high initial credences for a small number of possibilities and low credences for the other hypotheses. In the simulation, the concentration parameter is randomly chosen from a normal distribution centered at 0.5, with a standard deviation of 0.1. The reason for restricting the concentration parameter to a limited range is to prevent having priors too close to the uniform prior or too concentrated and thereby avoid giving small domain agents unfair advantage or disadvantage.
This simulation can thus also be interpreted simply as a guessing game of coin biases with monetary rewards involved.
Here I provide a more detailed example of the scoring rule. Suppose again that the doctor has a cluster infection of 3 patients, and the probability of death for each patient in the case at time t is 0.6. If the agent’s diagnosis is correct, she gains \((1-\frac{0.6}{a})\times 3\) points. If the intervention is incorrect, the agent gains \((1-\frac{1.6}{b})\times 3\) points, with a and b being the parameters explained in footnote 13. If after 100 time steps, no intervention has been made, then the agent gets \((1-p_{100})\times 3\) points, where \(p_{100}\) is the patient’s probability of death at the 100th time step.
The normal distribution is truncated at 0 and 1.
In the simulations, since the stakes range from 1 to 5, \(f(s)=\frac{s-3}{2}\). This yields the desired transformation with \(f(1) = -1\) and \(f(5) = 1\).
When \(s=3\), the threshold is the default \(\tau \) regardless of sensitivity, since \(f(3) = 0\), as specified in footnote 18.
Adjustments are made for cases where the domain size may be larger than N\(_{hyp}\).
The code for the simulations can be found in Appendix C.
Situations in which SmallSens(1) agents do form outright beliefs and take actions include (a) when the true coin bias is either 0 or 1, (b) when only one hypothesis in the active domain is neither 0 nor 1 (c) their credences are automatically rounded up to 1 by the computer, which happens at around the \(18^{th}\) decimal place.
The bars are generated from data averaged over runs with N\(_{hyp} = 5, 6\) for easy problems and over runs with N\(_{hyp} = 14, 15\) for difficult problems.
A related observation not shown in this graph is that the performances of small domain strategies are less affected by difficulty than full domain strategies. However, since the more possible alternatives there are, the more likely the truth is outside of the most probable \(4\pm 2\) hypotheses, there is still a negative correlation between the average scores and difficulty for small domain agents.
The results do not render FullSens, specifically in high stake situations, a failed strategy tout court, since there might be situations where errors are so severely penalized that FullSens agents’ level of prudence in high-stake scenarios is warranted. This is not the case in the present setup.
This suggestion requires the further assumption that an agent knows the costs of type 1 and type 2 errors, as well as how the costs change with time.
References
Babic, B. (2019). A theory of epistemic risk. Philosophy of Science, 86(3), 522–550. https://doi.org/10.1086/703552.
Christensen, D. (1991). Clever bookies and coherent beliefs. Philosophical Review, 100(2), 229–247. https://doi.org/10.2307/2185301.
Clarke, R. (2013). Belief is credence one (in context). Philosophers’ Imprint, 13, 1–18.
Cowan, N. (2008). What are the differences between long-term, short-term, and working memory? In W. S. Sossin, J.-C. Lacaille, V. F. Castellucci, & S. Belleville (Eds.), Essence of Memory, volume 169 of Progress in Brain Research (pp. 323–338). Elsevier. https://doi.org/10.1016/S0079-6123(07)00020-9.
Cowan, N. (2010). The magical mystery four: How is working memory capacity limited, and why? Current Directions in Psychological Science, 19(1), 51–57.
Cowan, N., Morey, C. C., Chen, Z., Gilchrist, A. L., & Saults, J. S. (2008). Theory and measurement of working memory capacity limits. In B. H. Ross (Ed.), Advances in Research and Theory, volume 49 of Psychology of Learning and Motivation (Vol. 49, pp. 49–104). Academic Press.
Douven, I. (2020). The ecological rationality of explanatory reasoning. Studies in History and Philosophy of Science Part A, 79, 1–14. https://doi.org/10.1016/j.shpsa.2019.06.004.
Duckworth, R. A., Potticary, A. L., & Badyaev, A. V. (2018). On the origins of adaptive behavioral complexity: Developmental channeling of structural trade-offs. In M. Naguib, L. Barrett, S. D. Healy, J. Podos, L. W. Simmons, & M. Zuk (Eds.), Advances in the study of behavior of advances in the study of behavior (Vol. 50, pp. 1–36). Boca Raton: Academic Press.
Easwaran, K. (2016). Dr. Truthlove or: How i learned to stop worrying and love Bayesian probabilities*. Noûs, 50(4), 816–853.
Fantl, J., & McGrath, M. (2002). Evidence, pragmatics, and justification. Philosophical Review, 111(1), 67–94. https://doi.org/10.1215/00318108-111-1-67.
Foley, R. (1992). Working without a net: A study of egocentric epistemology. Oxford University Press.
Frankish, K. (2009). Partial belief and flat-out belief (pp. 75–93). Springer.
Friedman, J. (2019). Inquiry and belief. Noûs, 53(2), 296–315. https://doi.org/10.1111/nous.12222.
Ganson, D. (2008). Evidentialism and pragmatic constraints on outright belief. Philosophical Studies, 139(3), 441–458. https://doi.org/10.1007/s11098-007-9133-9.
Gao, J. (2019). Credal pragmatism. Philosophical Studies, 176(6), 1595–1617. https://doi.org/10.1007/s11098-018-1081-z.
Gao, J. (2021). Credal sensitivism: Threshold vs. credence-one. Inquiry: An Interdisciplinary Journal of Philosophy. https://doi.org/10.1080/0020174X.2021.1933744.
Greco, D. (2013). 4. Probability and prodigality. Oxford Studies in Epistemology, 4, 82.
Greco, D. (2015). How I learned to stop worrying and love probability 1. Philosophical Perspectives, 29(1), 179–201. https://doi.org/10.1111/phpe.12059.
Hawthorne, J. (2003). Knowledge and lotteries. Oxford University Press.
Hawthorne, J., & Stanley, J. (2008). Knowledge and action. Journal of Philosophy, 105(10), 571–590. https://doi.org/10.5840/jphil20081051022.
Holton, R. (2014). Intention as a model for belief. In M. Vargas & G. Yaffe (Eds.), Rational and social agency: Essays on the philosophy of Michael Bratman. Oxford University Press.
Kyburg, H. E. (1961). Probability and the logic of rational belief. Wesleyan University Press.
Lee, M. D., Newell, B., & Vandekerckhove, J. (2014). Modeling the adaptation of search termination in human decision making. Decision, 1(4), 223–251.
Maher, P. (1993). The concept of acceptance. Cambridge studies in probability, induction and decision theory (pp. 130–161). Cambridge University Press. https://doi.org/10.1017/CBO9780511527326.007.
Mayseless, O., & Kruglanski, A. W. (1987). What makes you so sure? Effects of epistemic motivations on judgmental confidence. Organizational Behavior and Human Decision Processes, 39(2), 162–183. https://doi.org/10.1016/0749-5978(87)90036-7.
Miller, G A George A., Galanter, E., & Pribram, K. H. (1960). Plans and the structure of behavior. Holt, Rinehart and Winston.
Nagel, J. (2010). Epistemic anxiety and adaptive invariantism. Philosophical Perspectives, 24(1), 407–435. https://doi.org/10.1111/j.1520-8583.2010.00198.x.
Newell, B. R. (2005). Re-visions of rationality? Trends in Cognitive Sciences, 9(1), 11–15. https://doi.org/10.1016/j.tics.2004.11.005.
Norby, A. (2015). Uncertainty without all the doubt. Mind and Language, 30(1), 70–94. https://doi.org/10.1111/mila.12072.
Ross, J., & Schroeder, M. (2014). Belief, credence, and pragmatic encroachment. Philosophy and Phenomenological Research, 88(2), 259–288. https://doi.org/10.1111/j.1933-1592.2011.00552.x.
Rubin, K. (2015). Total pragmatic encroachment and epistemic permissiveness. Pacific Philosophical Quarterly, 96(1), 12–38. https://doi.org/10.1111/papq.12060.
Staffel, J. (2019). How do beliefs simplify reasoning? Noûs, 53(4), 937–962. https://doi.org/10.1111/nous.12254.
Stanley, J. (2005). Knowledge and practical interests. Oxford University Press.
Stanley, J. (2007). Replies to Gilbert Harman, ram neta, and Stephen Schiffer. Philosophy and Phenomenological Research, 75(1), 196–210. https://doi.org/10.1111/j.1933-1592.2007.00071.x.
Weatherson, B. (2005). Can we do without pragmatic encroachment. Philosophical Perspectives, 19(1), 417–443. https://doi.org/10.1111/j.1520-8583.2005.00068.x.
Weisberg, J. (2020). Belief in psyontology. Philosophers, 20(11), 1–8.
Williamson, T. (2000). Anti-luminosity (pp. 93–114). Oxford University Press.
Acknowledgements
I am grateful to Jonathan Weisberg, Jennifer Nagel, Boris Babic as well as two anonymous referees for their helpful comments on various drafts of this paper. Special thanks are also due to Titas Geryba for much-appreciated copy-editing and insightful suggestions. Thanks also to Nischal Mainali for interesting interdisciplinary discussions about cognitive constraints.
Funding
None.
Author information
Authors and Affiliations
Contributions
N/A.
Corresponding author
Ethics declarations
Conflict of interest
The authors declare that they have no conflict of interest.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendices
Appendix A: Weibull distribution
The Weibull distribution is a continuous probability distribution defined by the function
where \(\lambda \in (0, +\infty )\) is the scale parameter and \(k\in (0, +\infty )\) is the shape parameter. The cumulative distribution function of the Weibull distribution is given by
Appendix B: Symmetric dirichlet distribution
The Dirichlet distribution is a multivariate generalization of the beta distribution, parameterized by a vector \(\varvec{\alpha }\) with real entries. The Dirichlet distribution with parameters \(\alpha _1, \alpha _2, \ldots , \alpha _K\) is given by
where \(\sum _{i = 1}^{K} x_i = 1\) and \(x_1, x_2, \ldots , x_K > 0\). The normalizing constant \(B(\varvec{\alpha })\) is the multivariate beta function
where \(\Gamma \) is the gamma function.
A special case of the Dirichlet distribution is the symmetric Dirichlet Distribution. In a symmetric Dirichlet distribution, the entries of the vector \(\varvec{\alpha }\) are all equal. This corresponds to the case where there is no prior information that leads us to prefer one possibility over another. Since all the entries of \(\varvec{\alpha }\) are the same, the symmetric Dirichlet distribution can simply be parameterized by a scalar value \(\alpha \), often called the concentration parameter. When \(\alpha > 1\) the probability is randomly but evenly distributed across the possibilities. When \(\alpha < 1\), most of the probability will be concentrated in a few of the possibilities. The symmetric Dirichlet distribution is given by
Appendix C: Code for the model
The code for the model, written in Python using the agent-based simulation package Mesa, can be found here: https://github.com/alicecwhuang/threshold_code.git.
Rights and permissions
About this article
Cite this article
Huang, A.C.W. A normative comparison of threshold views through computer simulations. Synthese 200, 296 (2022). https://doi.org/10.1007/s11229-022-03784-x
Received:
Accepted:
Published:
DOI: https://doi.org/10.1007/s11229-022-03784-x