Trading Monotonicity for Cost in Beam Search

Sofia Lemons, Robert Holte, Carlos Linares López, Wheeler Ruml

Keywords: beam search, satisficing search

TL;DR: We introduce two new algorithms falling between regular beam search and monotonic beam search

Abstract: Beam search is a popular satisficing heuristic search algorithm, but increasing the beam width sometimes causes the algorithm to return a worse solution. A recent variant of beam search, monobead, guarantees nonincreasing solution cost with increasing beam width. However, the monotonicity of monobead sometimes comes at the price of increased cost and time for small beam widths. In this paper, we explore two algorithmic variants that lie between beam search and monobead. We find that, as hoped, our hybrids of beam and monobead can often find solutions with better cost than monobead and typically have more monotonic behavior than bead. This work improves our understanding of the price of monotonicity in beam search.