search

pettni / pdf-abstraction

Now developed in the mdp_network repository
1 stars 0 forks source link

Barrier certificate #6

Open shaesaert opened 5 years ago

shaesaert commented 5 years ago

The barrier certificates as developed by @pettni should be integrated into the code. Currently, the probability risk associated with taking an edge (or staying stationary and taking an observation) is given as a constant probability (.01). This effectively just discounts the value functions. The used constant is defined here: https://github.com/pettni/pdf-abstraction/blob/b232d56ee88793c1233568f9a2196240bfdc54a1/best/hVI_fsrm.py#L697 and it is used for discounting both the results of transition actions: https://github.com/pettni/pdf-abstraction/blob/b232d56ee88793c1233568f9a2196240bfdc54a1/best/hVI_fsrm.py#L713 and observation actions: https://github.com/pettni/pdf-abstraction/blob/b232d56ee88793c1233568f9a2196240bfdc54a1/best/hVI_fsrm.py#L754

pettni commented 5 years ago

I added a function (commit bf3de76) for computing things like this. I wasn't able to incorporate it into the FSRM since I couldn't understand how noise enters the dynamics. It seems to me like things are missing (Det_SI_Model seems deterministic (no noise) and SI_Model is not supported by the SPaths class). Also it seems like states are defined as points and not as sets with non-empty interior which is required for computations like these.

shaesaert commented 5 years ago

You are right, there is no noise introduced yet into the Det_SI_Model. I will work on it and see whether I can incorporate the barrier certificate function.

pettni commented 5 years ago

Epsilon-ball is fine, but epsilon probably shouldn't bee too small. I just pointed it out since I couldn't find any property giving the size or shape of states. It is not necessary that every point is associated to a discrete state (if that's what you meant).