Diverse Depth-First Search in Satisificing Planning

In satisficing planning where suboptimal plans are accepted, many planners use greedy best-first search (GBFS). Despite recent advances in automatic heuristic function generation, GBFS often suffers from performance degradation caused by inaccurate state evaluations. Diverse best-first search (DBFS) (Imai and Kishimoto 2011) avoids plateaus of search due to such inaccuracies by occasionally selecting states to expand that appear unpromising. Imai and Kishimoto showed that this approach outperforms the Fast Downward planner (Helmert 2006) with the best configurations based on GBFS for satisficing planning. Although DBFS has been shown to be effective, many hard planning problems remain unsolved. One drawback of DBFS is memory, beause DBFS is a best-first search algorithm and as such it must save all the open and closed states in memory, which can severely limit the scalability of DBFS.