An Efficient Real-Time Search Algorithm with Forecasting in Uncertain Problem Spaces

Real-time search is one of the most effective way when an agent can observe only limited information from its environment. RTA*, MTS, and their variations have been proposed as concrete algorithms for real-time search. However, if a heuristic value differs from a real value, an agent with these existing algorithms falls into the ”wrong” state whose heuristic value is small, and the agent might have difficulty reaching to a goal. In addition, because the existing real-time search algorithms have not considered the dynamic change of problem space, Efficiently solving the search problem in dynamic & uncertain problem spaces is difficult. In this paper, we propose a real-time search algorithm by forecasting to avoid falling into a state where the heuristic value is small. And we extend this algorithm to apply for uncertain problem spaces. We represent an uncertain problem space as a node-edge graph, and assume that the state of an edge becomes either changing to valid or invalid, and the agent knows only the probability of these changes of the edges. In this environment, to solve search problems efficiently, we propose a method that determines the action by considering the waiting time. In experiments, we investigated various situations on the complexity of problem space and on frequency of changes of environment. Our results demonstrate that our algorithm performs effectively in dynamic & uncertain environments and outperforms traditional real-time search algorithms.