Non-Clairvoyant Scheduling with Predictions

Permutation Predictions for Non-Clairvoyant Scheduling

“…It has several desirable properties such as (𝑖) monotonicity and (𝑖𝑖) Lipschitzness (both highly advertised recently by Im el al. [30]), (𝑖𝑖𝑖) theoretical learnability of our prediction model with respect to the error de nition, which we show by proving that our predictions are e ciently PAC-learnable in the agnostic sense, as well as (𝑖𝑣) practical learnability, which we demonstrate in empirical experiments, showing that our implemented learning algorithm quickly improves the performance of our scheduling algorithms and appears superior to previously presented algorithms.…”

“…In empirical experiments 1 we demonstrate the practicability of our approach in comparison to the previously proposed learning-augmented algorithms by Im et al [30] and Wei and Zhang [56]. These algorithms consider the single-machine problem without weights and release dates, 1|𝑝𝑚𝑡𝑛| 𝐶 𝑗 .…”

“…Im et al [30] advocate particularly the rst two properties. Monotonicity requires, in the length prediction model, that the error grows as more length predictions become incorrect.…”

“…In contrast to this example, there are other instances where 𝜈 and ℓ 1 underestimate the actual di culty that is caused by the inaccuracy of the prediction given to an (optimal) algorithm. Im et al [30] give such an example with 𝑝 1 = 𝑦 1 = . .…”

“…MultiStage (algorithm by Im et al [30]), works in phases and decides whether to follow the prediction or to execute RR by tracking the quality of the prediction. This is done by processing and computing the error of small random samples, which is then projected to the whole set of remaining jobs.…”