On Mon, Jan 23, 2017 at 7:59 AM, 'Nil Geisweiller' via opencog <opencog@googlegroups.com> wrote: > > The spread of the second order distribution shrinks as more evidence > accumulates, so it depends on the number of observations. There is a > function to translate the count N (number of observations) into confidence > > c = N / (N + K) > > so as you may see as the count increases, so does the confidence.
This is a rather weak sigmoid function, located at 'K'. I've often wondered how well inference chaining would work if one used a sharper formula. The above *eventually* approaches a confidence of 1.0, but it does it very slowly. Pathologically slowly, even. If I recall correctly, K is hard-coded as 800 in the code. What if, instead, one used c = tanh(N-K/K) which is a much cleaner, sharper sigmoid centered at K and at least one web page claims its very very fast, about as fast the simple formula (!!??) https://txt.arboreus.com/2013/03/29/fast-sigmoid.html I made a plot, see attached. The tanh formula in particular looks like it might give much better results for TV merging in the PLN chainer. --linas -- You received this message because you are subscribed to the Google Groups "opencog" group. To unsubscribe from this group and stop receiving emails from it, send an email to opencog+unsubscr...@googlegroups.com. To post to this group, send email to opencog@googlegroups.com. Visit this group at https://groups.google.com/group/opencog. To view this discussion on the web visit https://groups.google.com/d/msgid/opencog/CAHrUA36hbe3dpRXn3_gOXKmUKbKotajgGoyLhGO6nA0zBBu%3DNg%40mail.gmail.com. For more options, visit https://groups.google.com/d/optout.
sigmoid.gplot
Description: Binary data
