Bayesian updates
Each observation provides some information about the source location, which can be accounted for using Bayesian inference. In the case of a sequential process such as the source-tracking POMDP, Bayes’ rule can be applied after each observation to maintain an up-to-date belief \(p({\bf x})\) which encompasses all information gathered so far.
The update after observing \(o_t\) in \({\bf x}^a_{t+1}\) reads
where \(\text{Bayes}(p({\bf x}), {\bf x}^a, o)\) is the operator that maps the prior \(p_t\) to the posterior \(p_{t+1}\) through Bayes’s rule
and where \(\Pr(o | {\bf x}^a,{\bf x})\) is called the evidence in Bayesian terminology.
Let us now go through the update rule for each observation.
If \(o=F\), the source has been found in \({\bf x}^a\), and the posterior distribution is simply a Dirac distribution
Otherwise, \(o=(\bar{F}, h)\), meaning that the source has not been found and that \(h\) hits were perceived. The posterior distribution after not finding the source is a simple renormalization
The posterior after a hit \(h\) is
The full update after observing \(o=(\bar{F}, h)\) is therefore given by the successive application of each partial update
A step-by-step search shows how Bayesian updates are computed through an example.