opencog-dev team mailing list archive

["Joel Pitt" <joel.pitt@xxxxxxxxx>]

Hi Ben,

Quick question about using Markov matrices for Importance spread.

In the initial suggestion you wrote:
> I'm wondering if we could think about this in terms of Markov matrices.
>
> Let the variable p_i equal the "probability that Atom A_i is selected",
> which should be proportional to its STI
>
> Then the Hebbian and InverseHebbianLinks should determine the
> transition probabilities p_ij, right?   (The probability that A_i is selected,
> given that A_j was selected)
>
> Then, we have a Markov matrix M.
>
> We can find its inverse matrix N
>
> Perhaps we then want to multiply q=N p, where p=(p_1,...,p_n)
>
> ... and set the STI levels equal to the entries of q?

However, after brushing up on some linear algebra, it seems to me that
we don't need to find the inverse matrix at all.

Representing spread as a Markov process, each potential
source/destination atom is a state which a portion of STI can be in.
p_n represents the initial state (i.e. the distribution of STI across
atoms) and we merely have to multiply p by M to get the next state
p_(n+1)

Make sense?

J


The Markov or transition matrix is simply multiplied by the
distribution of importance across states.