Interval Selection with Binary Predictions

Abstract

Following a line of work that takes advantage of vast machine-learned data to enhance online algorithms with (possibly erroneous) information about future inputs, we consider predictions in the context of deterministic algorithms for the problem of selecting a maximum weight independent set of intervals arriving on the real line. We look at two weight functions, unit (constant) weights, and weights proportional to the interval's length. In the classical online model of irrevocable decisions, no algorithm can achieve constant competitiveness (Bachmann et al. [BHS13] for unit, Lipton and Tomkins [LT94] for proportional). In this setting, we show that a simple algorithm that is faithful to the predictions is optimal, and achieves an objective value of at least $OPT -\eta$, with $\eta$ being the total error in the predictions, both for unit, and proportional weights. When revocable acceptances (a form of preemption) are allowed, the optimal deterministic algorithm for unit weights is $2k$-competitive [BK23], where $k$ is the number of different interval lengths. We give an algorithm with performance $OPT - \eta$ (and therefore $1$-consistent), that is also $(2k +1)$-robust. For proportional weights, Garay et al. [GGKMY97] give an optimal $(2\phi + 1)$-competitive algorithm, where $\phi$ is the golden ratio. We present an algorithm with parameter $\lambda > 1$ that is $\frac{3\lambda}{\lambda -1}$-consistent, and $\frac{4\lambda^2 +2\lambda}{\lambda -1}$-robust. Although these bounds are not tight, we show that for $\lambda > 3.42$ we achieve consistency better than the optimal online guarantee in [GGKMY97], while maintaining bounded robustness. We conclude with some experimental results on real-world data that complement our theoretical findings, and show the benefit of prediction algorithms for online interval selection, even in the presence of high error.