[Paper Review] Efficiency of a Stochastic Search with Punctual and Costly Restarts

Kabir Husain, Sandeep Krishna|arXiv (Cornell University)|Sep 13, 2016

Optimization and Search Problems3 citations

TL;DR

This paper develops a general analytical framework to optimize stochastic search processes with arbitrary restart protocols, using path enumeration and Laplace transforms to compute joint distributions of completion time and restart count. It shows that optimal restart times scale linearly with restart distribution variance and identifies trade-offs between speed, cost, and restart frequency, with deterministic restarts outperforming stochastic ones under certain conditions.

ABSTRACT

The mean completion time of a stochastic process may be rendered finite and minimised by a judiciously chosen restart protocol, which may either be stochastic or deterministic. Here we study analytically an arbitrary stochastic search subject to an arbitrary restart protocol, each characterised by a distribution of waiting times. By a direct enumeration of paths we construct the joint distribution of completion time and restart number, in a form amenable to analytical evaluation or quadrature; thereby we optimise the search over both time and potentially costly restart events. Analysing the effect of a punctual, i.e. almost deterministic, restart, we demonstrate that the optimal completion time always increases proportionately with the variance of the restart distribution; the constant of proportionality depends only on the search process. We go on to establish simple bounds on the optimal restart time. Our results are relevant to the analysis and rational design of efficient and optimal restart protocols.

Motivation & Objective

• To develop a general analytical method for optimizing stochastic search processes with arbitrary restart protocols.
• To analyze the impact of stochasticity in restart timing on mean completion time and optimality.
• To incorporate restart costs into the optimization framework, moving beyond minimization of mean completion time alone.
• To derive analytical bounds on the optimal restart time for various search processes.
• To compare deterministic and stochastic restart protocols under both time and cost constraints.

Proposed method

• Uses direct path enumeration to compute the joint probability distribution of completion time and number of restarts.
• Applies Laplace transforms to the joint distribution to enable analytical evaluation of moments and cost functions.
• Models the search process as a three-state system (A → B → C, with B → A as restart), with arbitrary waiting time distributions for each transition.
• Derives expressions for the mean completion time and mean number of restarts via generating functions and the Hurwitz-Lerch transcendent.
• Introduces a cost function linear in restart count and generalizes it to include time-cost trade-offs (e.g., $ f(m,T) = m^eta T $).
• Uses quadrature and special functions (e.g., $ ilde{ heta}(s) $, $ ilde{ ho}(s) $, $ ilde{ u}(s) $) to evaluate moments and optimize protocols.

Experimental results

Research questions

• RQ1 How does the variance of a restart distribution affect the optimal mean completion time in a stochastic search?
• RQ2 What are the analytical bounds on the optimal deterministic restart time for a given search process?
• RQ3 How do deterministic and stochastic restart protocols compare in terms of efficiency and cost when each restart incurs a fixed cost?
• RQ4 Can a unified cost function combining completion time and restart count be optimized analytically for arbitrary restart and search processes?
• RQ5 Under what conditions does deterministic restarting remain optimal when restarts are costly?

Key findings

• The optimal completion time increases linearly with the variance of the restart distribution, with the proportionality constant depending only on the search process.
• The mode of the search distribution $ P_s(t) $ provides a strict lower bound for the optimal deterministic restart time $ au_{ ext{opt}} $, i.e., $ au_{ ext{opt}} > ext{mode}(P_s) $.
• For 1D diffusion with Poisson restarts, deterministic restarts incur lower average restart costs than stochastic ones when the restart rate $ r < 1.412 $, but exceed them at higher rates.
• The cost function $ ig angle m^eta T ig ext{ for } eta > 0 $ shows that deterministic restarting remains more optimal across various time-cost trade-offs.
• The framework enables analytical computation of moments and cost functions via quadrature, reducing complex path integrals to manageable special functions.
• The method is general and applies to arbitrary search, restart, and delay distributions, including non-Poissonian and non-exponential restart processes.