Beam Search: Faster and Monotonic

Abstract

Beam search is a popular satisficing approach to heuristic search problems that allows one to trade increased computation time for lower solution cost by increasing the beam width parameter. We make two contributions to the study of beam search. First, we show how to make beam search monotonic; that is, we provide a new variant that guarantees nonincreasing solution cost as the beam width is increased. This makes setting the beam parameter much easier. Second, we show how using distance-to-go estimates can allow beam search to find better solutions more quickly in domains with non-uniform costs. Together, these results improve the practical effectiveness of beam search.

Tunable Suboptimal Heuristic Search

“…We find through experimental evaluation that Speed* often outperforms not just wA* but also RR-d, the state-of-the-art bounded suboptimal algorithm. We also see that Speed* is robust to domains with dead-ends, unlike beam-based search approaches such as Bead (Lemons et al 2022) and Rectangle search (Lemons et al 2024). This work is the first to evaluate RR-d and Rectangle in the tunable suboptimal setting.…”

“…Rectangle is a state-of-the-art anytime algorithm based on beam search (Lemons et al 2024). Rectangle can be thought of as a beam search whose width increases as the search tree depth increases, according to an 'aspect ratio' specified at runtime.…”

Rectangle Search: An Anytime Beam Search

“…It continues to search until either a solution is found or until no new states are reachable from the current depth. A variant of beam search using d(n) for node selection called bead outperforms beam search using f (n) or h(n) in non-unit cost domains (Lemons et al 2022).…”

“…One issue with beam search is that when the beam width b is increased, sometimes a higher cost solution is returned. The algorithm monobead (Lemons et al 2022) addresses this issue. It regards the beam as an ordered sequence of numbered slots.…”

“…To keep our plots clear, we only include the competitive algorithms: ARA*, AEES, CABS, and rectangle search. Plots for all algorithms in all domains are given in Lemons et al (2023). We ran ARA* in several previously-proposed configurations: initial weights of 10 and 2.5 with a decrement of 0.02 (Likhachev, Gordon, and Thrun 2004) and a weight schedule of 5, 3, 2, 1.5, 1 (Thayer, Benton, and Helmert 2012).…”