Utilisateur:JP Garcia Ballester

Un article de APMC.

Sommaire

My current work on APMC

As you surely know, APMC calculates probabilities of LTL formulas. It means you give it a formula φ, a precision parameter ε, a confidence parameter δ, and APMC calculates ρ, an approximation of the probability p of φ being true.

When testing p against a threshold θ, we do not always need a high precision. Indeed, when p is far from θ, the precision required is lower than when p is near θ. We just need to be sure that p is superior or inferior to θ. For example, suppose p is 0.7 and we want to know whether it is superior or inferior to 0.5: if we calculate ρ to be 0.8 with ε equals 0.2, we do not need a more accurate precision.

Therefore, I'm working on an iterative version of APMC which would stop when ε is big enough to take a decision. We know

p > ρ - ε
p < ρ + ε

Therefore, decision can be reach when ρ - ε > θ or ρ + ε < θ.

Here is the algorithm doing this optimization:

Input: diagram, φ, θ, δ, k
A := 0
N := 0
R := 0
 While R = 0 do
 Generate a random path σ of length k with the diagram
 If φ is true on σ then A := A + 1
 N := N + 1, ε := sqrt(4log(2/δ)/N), ρ := A / N
 If ρ - ε > θ then R := 1
 If ρ + ε < θ then R := -1

This algorithm has been implemented in the standalone Java version and in the cluster implementation of APMC.

Some unuseful informations about myself

Some nearly less unuseful informations about myself

  • My mail: giga@le-pec.org
  • My jabber-id: giga@le-pec.org
  • My phone number: (+33) 06 31 78 40 48

My other projects

  • MouliRuby: Moulette written in ruby
  • LibSSH: A SSH library in C. I only take care of libgcrypt port