Re: Decyphering lies (was: Re: Anthrax addendum.)y

From: Anders Sandberg (
Date: Mon Oct 15 2001 - 08:04:02 MDT

On Mon, Oct 15, 2001 at 09:12:18AM -0400, Alex F. Bokov wrote:
> This raises an interesting question: how to form a rational opinion
> when most sources of information claim to be unbiased and all of them
> are biased? The US goverment lies. The enemies of the US government
> lie. The mass media is owned by like a dozen people, and what are the
> chances that they're impartial in all this? I wonder if game theory or
> information theory have some strategies to offer in juxtaposing the
> 'information' coming from different sources so the opposing lies
> cancel each other out somewhat?

A very good question. My personal strategy is to have as many new
sources as possible with as different biases as possible, comparing them
with each other. Sure, both CNN and Al Jazira are biased, but not in the
same way - and the bias of Slashdot is very different from the bias of
Dagens Nyheter.

Let's see if we can do some information theory here. Assume we have an
information source S (the Truth), which we get information about through
noisy and biased channels. In some cases we know the bias, in some cases
it is unknown. What is the optimal estimate of S?

To make things more concrete, assume S is a scalar we want to estimate
(like the probability of Bin Laden being a nice guy or the number of
dead in a bombing). From information channel i we receive an estimate
s_i, which is s_i = a_i S + N(b_i,c_i): - a_i is how much S is
distorted, b_i is the bias and c_i the standard deviation. The
distribution of s_i becomes f(s_i | S) = N(x-a_iS-b_i,c_i). Here I
assume independent channels (BIG assumption!).

Given data from s_i, we can then do a maximum likelyhood estimation of
a_i, b_i and c_i if we hold S constant (i.e compare the news about a
certain fact for a while).

Bayes rule says that the distribution of S given s_i is f(S | s_1, s_2,
...) = [integral f(S,s) ds] f(s | S) / integral [integral f(S,s) ds] f(s
| S) dS where f(S,s) is the joint distribution of S and s_1, s_2,.... We
already have f(s_i|S), so f(s|S)= product f(s_i|S). So we get f(S| s) =
f(S) product f(s_i|S) / <normalizing stuff>.

We still have the Bayes factor f(S) - we need an estimate of that. This
is where our own preconceptions and world assumptions come in. Usually I
think we simply assume it is normally distributed and rescaled to have
zero mean and 1 variance.

OK, this seems to suggest a method: gather s_i, for each channel create
a maximum likeliehood estimate of a_i, b_i and c_i given the assumption
S=N(0,1), put it into the formula above and we get a
posterior distribution of S.

If channels are of uncertain independence, we can either calculate
correlations between them for a while, and then select a set that are
uncorrelated. A more clever method would be to use principal component
analysis to create "virtual channels" that are weighted sums of the
channels that are mutually independent and hence could be used to
estimate S.

Caveat: I'm not a staticistician and this was just off the cuff - there
has to be some serious analysis on this, I just don't know where to
look (and I don't have the time).

> I only have a crude approximation of such a strategy, where I assign
> higher credibility to an institution if liberals accuse it of being
> conservative and conservatives accuse it of being liberal. There are
> many domains this heuristic doesn't cover though, so I'm looking for
> others.

A nice strategy, I like it!

This is a bit like finding a channel that is independent of two known
anticorrelated channels.

Anders Sandberg                                      Towards Ascension!                  
GCS/M/S/O d++ -p+ c++++ !l u+ e++ m++ s+/+ n--- h+/* f+ g+ w++ t+ r+ !y

This archive was generated by hypermail 2b30 : Sat May 11 2002 - 17:44:13 MDT