Evaluation of Different Algorithms on the RDDL Simulation Platform

Probabilistic and Parallel Planning (PPP) is a hot spot in AI planning in recent years. Relational Dynamic Influence Diagram Language (RDDL) is a new language used in PPP. For PPP problems, Real-Time Dynamic Programming (RTDP) and Upper Confidence Bounds (UCB1) Applied to Trees (UCT) algorithm are few approaches to find near-optimal solutions. However, there is no evaluation of the performance of these algorithms under the same platform. Different planners based on RTDP or UCT have different performance. It is hard to answer which algorithm is better. Fortunately, RDDLSim is a platform which implements a parser, simulator, and client/server evaluation architecture for RDDL. Empirically testing these algorithms on the same platform is necessary. In this paper, we evaluate the performance of RTDP and UCT in RDDLSim. Experimental results show that the performances of these two algorithms are domain specific. In several domains, RTDP is significantly more efficient than its opponent. While in other domains, UCT performs better.