Towards real-time search with inadmissible heuristics

Real-time search has two aspects, one as an efficient search method (in a single problem solving trial), and the other as an overall problem solving architecture with learning ability (through repeated trials). In both respects, the use of inadmissible (and hence inconsistent) heuristic functions bring some merits such as improved performance, but there is no theory yet that well explains when and why these algorithms benefit from them. In this paper, as a step towards fully understand and take advantage of such nonstandard heuristic functions, we discuss the properties of LRTA* and the Moving-Target Search (MTS) algorithms under heuristic functions violating admissibility or consistency. In particular, we show (1) the completeness of MTS with inconsistent or even inadmissible heuristics, and present (2) a new proof technique for the convergence of LRTA* which is applicable regardless of consistency.