A Monte Carlo Update for Parametric POMDPs

This paper presents the Parameterised POMDP (PPOMDP) algorithm: a method for planning in the space of continuous parameterised functions. The novel contribution is an approach to transitioning parameterised beliefs using Monte Carlo methods. By re-using prediction and observation calculations, the transition function can be computed efficiently. An analysis of scalability suggests that the approach is likely to scale to physically larger environments than algorithms which rely on an underlying discretisation. Experimental results in a simulated robot navigation problem show that the algorithm compares favourably with existing approaches.