An Incremental Search Approach to Real-Time Decision Making

Decision-Theoretic Planning

Volume

Papers from the 1994 AAAI Spring Symposium

Contents

We propose incremental, real-time search as a general approach to real-time decision making. A good example of real-time decision making is air-traffic control when an airport is unexpectedly closed due to an earthquake or severe winter storm. Under these circumstances, the controller must decide in real time where to divert each incoming plane so that it can land safely. We model real-time decision making as incremental tree search with a limited number of node expansions between decisions. We show that the decision policy of moving toward the best frontier node is not optimal, but nevertheless performs nearly as well as an expected-valuebased decision policy. We also show that the real-time constraint causes difficulties for traditional best-first search algorithms. We then present a new algorithm that uses a separate heuristic function for choosing where to explore and which decision to make. Empirical results for random trees show that our new algorithm outperforms the traditional best-first search approach to real-time decision making. Our results also show that depth-first branch-and-bound performs nearly as well as the more complicated best-first variation.