An Experimental Comparison of Classical, FOND and Probabilistic Planning

Abstract

Abstract. Domain-independent planning in general is broadly applicable to a wide range of tasks. Many formalisms exist that allow the description of different aspects of realistic problems. Which one to use is often no obvious choice, since a higher degree of expressiveness usually comes with an increased planning time and/or a decreased policy quality. Under the assumption that hard guarantees are not required, users are faced with a decision between multiple approaches. As a generic model we use a probabilistic description in the form of Markov Decision Processes (MDPs). We define abstracting translations into a classical planning formalism and fully observable nondeterministic planning. Our goal is to give insight into how stateof-the-art systems perform on different MDP planning domains.

From FOND to Robust Probabilistic Planning: Computing Compact Policies that Bypass Avoidable Deadends

“…We are not alone in recognizing the computational core that is shared by FOND and probabilistic planning (cf. (Hertle et al 2014)). With this work, we have demonstrated the merit of this correspondence by exploiting compact policy representations, relevance reasoning, and deadend avoidance techniques developed within the FOND community, and used these to advance the state of the art in probabilistic planning.…”

“…It is well-known that solutions to the FOND problem FOND(P) that result from ignoring transition probabilities in a probabilistic planning problem P are also solutions to P. Moreover, strong-cyclic solutions to FOND(P) are optimal High-Prob solutions to P and vice-versa (cf. (Hertle et al 2014)).…”

“…Recent advances in FOND planning have resulted in highly optimized planners that scale well, computing compact policies offline that bypass avoidable deadends (e.g., (Muise, McIlraith, and Beck 2012)). Our interest is in exploiting the computational core shared by FOND and probabilistic planning, as noted in (Hertle et al 2014), to investigate whether state-of-the-art FOND planning techniques can be leveraged to produce an offline probabilistic planner that also bypasses avoidable deadends, scales well, and produces compact policies. We focus on what we will refer to as the Copyright c 2016, Association for the Advancement of Artificial Intelligence (www.aaai.org).…”

On-Line Case-Based Policy Learning for Automated Planning in Probabilistic Environments

“…In fact, deterministic planning is the most preferable option used in execution-monitoring-planning loops. Deterministic planners has advantages in computation time and execution efficiency, while probabilistic planners tend to produce overly cautious and thus overly long plans 12,27 tionally, in real-world problems the transition model (i.e., the probabilities of the actions' effects) are unknown and replanners can deal with these problems without a transition model and without a random exploration of the state space in search of better and better policies 3,1 .…”