Decision-theoretic planning via probabilistic programming
We study planning in Markov decision processes involving discrete and continuous states and actions, and an unknown number of objects. Planning in such domains is notoriously challenging and often requires restrictive assumptions. We introduce HYPE: a sample-based planner for hybrid domains that is very general, which combines model-based approaches with state abstraction. Most significantly, the domains where such planners are deployed are usually very complex with deep structural and geometric constraints. HYPE is instantiated in a probabilistic programming language that allows compact codification of such constraints.
In our empirical evaluations, we show that HYPE is a general and widely applicable planner in domains ranging from strictly discrete to strictly continuous to hybrid ones. Moreover, empirical results showed that abstraction provides significant improvements.
In the final part of the talk, we turn to the question of whether there is any hope of developing computational methodologies that are not based on sampling. In particular, it is tricky in hybrid domains to deal with low-probability observations, and most sampling-based schemes only provide asymptotic guarantees.
This talk is based on a Machine Learning Journal article (2017), and is joint work with Davide Nitti, Tinne De Laet and Luc De Raedt.