Improvements to the Strong Cyclic Planning Algorithm

Abstract

This paper points out some drawbacks and gives some modifications to the strong cyclic planning algorithm. In particular, a set of states is found by use of adjacency matrix that corresponds to a non-deterministic state-transition system. The set of states is composed of two parts, the first part is the states that are unreachable from the initial states, so the station-action pairs which relate to these states are absolutely useless; the second part is the states which the initial states can not reach without passing any goal state before, the state-action pairs which relate to the second part are useless as well, because they moves the execution away from the goal. So a great many of state-action pairs can be eliminated directly from the universal policy. In addition, there are many repeated computations in function RemoveNonProgress, function WeakPrelmage adds some state-action pairs into Prelmage, however, many of these state-action pairs are eliminated by function PruneStates later. In the modified algorithm, we design a new function to avoid these repeated computations. Finally, the efficiency of the modified algorithm is illustrated by an example and some experiments.