Anytime Heuristic Search: Frameworks and Algorithms

Abstract

Anytime search is a pragmatic approach for trading solution cost and solving time. It can also be used for solving problems within a time bound. Three frameworks for constructing anytime algorithms from bounded suboptimal search have been proposed: continuing search, repairing search, and restarting search, but what combination of suboptimal search and anytime framework performs best? An extensive empirical evaluation results in several novel algorithms and reveals that the relative performance of frameworks is essentially fixed, with the repairing framework having the strongest overall performance. As part of our study, we present two enhancements to Anytime Window A* that allow it to solve a wider range of problems and hastens its convergance on optimal solutions.

Better Parameter-Free Anytime Search by Minimizing Time Between Solutions

“…The largest drawback of the continued framework is that it does not reconsider its parameters as new incumbents are found. The search strategy used to find the first solution is the same as that used to find the last solution, and this can be inefficient (Thayer and Ruml 2010). Repairing Search addresses this shortcoming of the continued search framework, and it's designed to work well in domains with many duplicates (Likhachev, Gordon, and Thrun 2003).…”

“…We use this dynamic bound to set w for the next iteration of AEES. This technique has also been used to augment parameter schedules used by anytime search (Likhachev, Gordon, and Thrun 2003;Hansen and Zhou 2007;Thayer and Ruml 2010).…”

Deadline-Aware Search Using On-Line Measures of Behavior

“…Depthfirst branch-and-bound is a well-known anytime algorithm, but it applies only to problems in which duplicate states are rare and a useful bound is known on the depth of solutions, criteria which exclude many domains, including the benchmarks considered here. Thayer and Ruml (2010) performed an evaluation of several anytime methods across a wide variety of domains. The repairing framework of Anytime Repairing A* (Likhachev, Gordon, and Thrun (2003)) had the best general performance across all domains and search strategies, making it the prime anytime competitor for deadline search.…”

“…Heuristics and g(s) values are cached between iterations, while the open list is cleared. While RWA* is generally slower than ARA* (Thayer and Ruml 2010), in some search domains, particularly domain independent planning, restarting the search at each iteration has a surprisingly positive effect on the results. We include results for this algorithm as well as ARA* in our evaluation.…”

To Reopen or Not To Reopen in the Context of Weighted A*. Classifications of Different Trends (Extended Abstract)

“…The notion of reopening has been first introduced by (Pohl 1970) and discussed in a variety of papers thereafter (Likhachev, Gordon, and Thrun 2003;Hansen and Zhou 2007;Thayer and Ruml 2008). A number of authors discussed the influence of AR and of NR on the path returned and on the number of expanded nodes (Thayer and Ruml 2010;Malima and Sabanovic 2007;Valenzano, Sturtevant, and Schaeffer 2014). For example, (Hansen and Zhou 2007) discuss some reasons for why AR may require fewer node expansions than NR in some cases, but did not discuss the case where AR finds a worse solutions than NR.…”