(half-baked brainstorming below, beware....  What I'm musing about is
how to guess the causal direction of a correlation based on
non-temporal data...)

> I've been re-reading this nice old paper on the foundations of the Second 
> Law..
>
> http://necsi.edu/projects/baranger/cce.pdf

It's a physics-y paper but I think one can apply it to AGI with some
appropriate set-up

The key thing that Baranger's arguments show there is that --
Within the view of a coarse-graining observer (one whose precision of
observation is much less than the precision of the universe he's
observing), it's more likely for

-- two states that seemed the same at time T, to seem different at time T+1

than for

-- two states that seemed different at time T, to seem the same at time T+1

(this is for an arbitrary trajectory in a conservative dynamical
system, blabla...)

Now, suppose we apply this reasoning (hands waving kinda wildly) to a
space of **situations** in some universe.  Each point in the state
space is a certain situation.   A trajectory in the state space is a
series of situations, e.g. the series of situations encountered by
some agent.  Suppose that the trajectories of situations encountered
by agents, when plotted in situation-space, are complex and
fractal-looking like the ones in Baranger's paper.  Each agent may be
associated with a probability distribution over trajectories (the
possible histories it experiences).

A possible commonsensical cause or effect like "rain" or "dark", in
this framework, corresponds to a set of situations (e.g. the
situations involving rain).   Thus it corresponds to a certain region
in the situation space.  Let's call these "event-sets".   Each point
on a trajectory through situation-space is going to pass through
various event-sets.

To say what it means for one event-set to cause another, relative to a
certain set of trajectories (or probability distribution over
trajectories), we can use the definitions from Luke Glynn's paper
http://philsci-archive.pitt.edu/9729/1/Website_Version_2.pdf

What Baranger's line of argument (via which he derives the Second Law)
suggests is that overall

-- same cause, different effects

is more likely than

-- different cause, same effects

This is because "same cause, different effects" means "two different
situations, which are put into the same event-set by the observer,
lead to two different situations, which are put into different
event-sets by the observer", etc.

Since event-sets are regions of situation-space, and generally
(because of the coarse-graining observer) an earlier time-point on a
trajectory is going to be less spread-out through situation-space than
a later time-point on the same trajectory --- therefore the cause is
likely to be less spread-out than the effect.

Thus overall we might conclude: given a pair  of event-sets (X,Y) that
are correlated (meaning e.g. that there is mutual information between
the distribution of particular events within category X, and the
distribution of particular events within category Y),

--  the one with greater spread (i.e. the greatest differentiation,
i.e. the greatest entropy, among the different particular situations
in the event-set) is more likely to be in the future...

The basic idea is: if event-categories X and event-categories Y are
sufficiently correlated that it seems likely one of

A)  The states of the universe corresponding to observation of X tend
to causally affect the states of the universe corresponding to
observation of Y [within the assumed set of trajectories along which
causation is being estimated]

or

B) The states of the universe corresponding to observation of Y tend
to causally affect the states of the universe corresponding to
observation of X [within the assumed set of trajectories along which
causation is being estimated]

then, to choose between X and Y, on the average we will guess right
more often if we assume the lower-entropy one of X and Y is the cause
and the higher-entropy one is the effect ...

So according to this way of thinking, the asymmetry required in
Glynn's analysis of causality could potentially be taken as entropy
rather than time...

maybe ;)

-- Ben


-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424
Modify Your Subscription: 
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657
Powered by Listbox: http://www.listbox.com

Reply via email to