search

pettni / pdf-abstraction

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

Resampling & pruning belief points #7

Open shaesaert opened 6 years ago

shaesaert commented 6 years ago

The belief points are currently only selected at the initialization of the algorithm. As a next step, the belief points should be updated during the iterations. This is, in general, a crucial point for point-based value iteration.

shaesaert commented 6 years ago

@rohant4255 perhaps this is the right source to look for an algorithm for the pruning and adding of belief points? https://arxiv.org/abs/1207.1412

shaesaert commented 6 years ago

The method proposed by Smith in "Point-based POMDP Algorithms: Improved Analysis and Implementation. " relies on the computation of an upper-bound. Additionally, they say it is only worth it for very very large POMDPs.

The alternative is to use the basic method, as explained in Pineau 2003; "Point-based value iteration: An Anytime algorithm for POMDPs." (http://www.fore.robot.cc/papers/Pineau03a.pdf) Sec.3.2. Belief point set expansion: They stochastically simulate a single step forward using each of the possible actions and keep the beliefs with the largest distance to the other beliefs. This should be easy to implement.

shaesaert commented 5 years ago

Sampling of belief points has been implemented in the current version.