AI Solves Complex Scheduling with Near-Optimal Guarantees

Scheduling tasks efficiently on a single machine is a fundamental problem in computer science with wide-ranging applications, from managing server workloads to optimizing production lines. Traditionally, this involves balancing job processing times, release times, and varying costs based on completion times, where some jobs are urgent and others can tolerate delays. The General Scheduling Problem (GSP) encapsulates this complexity by allowing each job to have its own cost function, making it a unifying framework for many well-studied scenarios like minimizing weighted flow time or tardiness. Until now, the best known result was a polynomial-time algorithm with an approximation ratio that scales logarithmically with processing time ratios, leaving a gap in achieving constant-factor guarantees. Researchers have now developed a quasi-polynomial time algorithm that provides a (2 + ε)-approximation for GSP, assuming processing times are quasi-polynomially bounded integers, marking a significant step toward practical near-optimal solutions.

This breakthrough stems from a reduction of GSP to a geometric covering problem called the Rectangle Covering Problem (RCP), where jobs are represented as rectangles in a plane and demands as rays. In RCP, the goal is to select a subset of non-overlapping rectangles, organized in rows, to cover the demand of vertical rays while minimizing total cost, with the constraint that from each row, only a prefix of rectangles can be chosen. The reduction loses only a factor of 1 + ε in objective value, transforming the scheduling into a more tractable geometric form. Previous work for special cases like weighted flow time leveraged a tree-like structure in the rectangles to guide algorithms, but GSP's arbitrary cost functions prevent such simplifications, necessitating a novel approach to handle the lack of inherent structure.

The algorithm for RCP employs a recursive strategy that splits the problem area vertically into left and right subproblems, handling rows that cross the split line—centered, right-sticking-in, left-sticking-in, and spanning rows—differently to avoid accumulating approximation errors. Key to is transforming an optimal solution into a structured near-optimal one with additional rectangles, creating slack that allows guessing parts of the solution efficiently. For instance, in rows where many rectangles are selected, the algorithm adds extra rectangles from bottom rows, oversatisfying some ray demands and enabling polylogarithmic guessing of artificial rays. This structured approach ensures that the approximation ratio does not degrade with recursion depth, ultimately achieving a (2 + ε)-factor for the general case and an improved (1 + ε)-factor for the weighted tardiness special case.

From the paper show that for GSP, the algorithm runs in quasi-polynomial time, specifically 2^poly((1/ε) log(n+pmax))^{1/ε}, where n is the number of jobs and pmax is the maximum processing time. For weighted tardiness, the approximation ratio improves to 1 + ε with the same time complexity, leveraging a rounding technique that discretizes rectangle coordinates to exploit weaker structural properties. The reduction to RCP ensures that the number of rectangles per row is bounded, and the algorithm's recursive partitioning handles up to O(log T) levels, with each step involving guessing over a limited number of groups and rows. Empirical include potential applications in logistics, where jobs with heterogeneous deadlines and priorities can be scheduled with near-minimal delay costs, enhancing efficiency in supply chains and data centers.

Despite these advances, limitations remain. The algorithm assumes quasi-polynomially bounded integer processing times, which may not hold in all practical scenarios with unbounded data. Additionally, the quasi-polynomial running time, while an improvement over exponential approaches, may still be prohibitive for very large instances, though it represents progress toward polynomial-time constant approximations. The paper also notes that for weighted tardiness, the result implies the problem cannot be APX-hard under certain complexity assumptions, but open questions persist about achieving a polynomial-time approximation scheme (PTAS) for GSP. Future work could explore extensions to multiple machines or online settings, where real-time decision-making is crucial.