Hi, Playing around with the classical problem of inferring the bias of a coin based on observed outcomes, I noticed the strange behavior for extreme observations. Imagine we toss 30 times a coin and we get only heads (ones). We model the problem as a Bernoulli distribution with a flat prior for p …

Agreed with @ricardoV94 , that you are seeing beta(2,1). Not sure why (it’s as if np.ones(30) is being treated as if it’s np.array([1]) instead), but if you wrap your numpy array in a pm.Data object, it seems to recognize all 30 flips, even when they are all heads: with pm.Model() as coin_flipping:…

Bernouilli corner case, all heads coin?

Questions

ricardoV94 December 26, 2020, 9:13pm 2

Something about your first plot / trace is definitely wrong. It looks exactly like a beta(2,1), and not the beta(31,1) it should look like.

Topic		Replies	Views
Modelling three biased coins from generated data Questions	0	358	April 4, 2020
Revisiting the coin-flipping problem Questions	6	3517	July 20, 2019
Parameter Learning in Simple Bayesian Network Questions	5	2828	April 2, 2018
The results of a switch are observed directly (beginner) Questions	1	924	November 11, 2018
Beta Binomial distribution Questions	1	700	August 26, 2018

Bernouilli corner case, all heads coin?

Related topics