Decision-making agents are often faced with uncertainty about their current state or their environment. Partially Observable Markov Decision Processes (POMDPs) are a well-established model for dealing with uncertainty and stochastic behaviour. Most work focuses on analyzing given POMDPs. However, a system designer may have at least limited control over what an agent can observe about its environment, e.g. by choosing the available sensors. We study the problem of determining optimal (wrt. helping or hindering an agent in achieving its goals) observability capabilities of a POMDP under various constraints such as budget limitations or preselected observables. We show how optimal strategies of the underlying fully-observable MDP can be used to derive effective POMDPs and their strategies. We evaluate our algorithms with the probabilistic model checker PRISM.
Join at: imt.lu/sagrestia