This paper explores how machine learning (ML) predictions can be used to improve the performance of online algorithms. The authors study two classical online problems: ski rental and non-clairvoyant job scheduling. In the ski rental problem, a skier must decide whether to rent or buy skis based on the number of days they will ski. In non-clairvoyant job scheduling, the goal is to schedule jobs on a single machine with unknown processing times. The authors propose new online algorithms that use ML predictions to make decisions, achieving better performance than traditional algorithms while maintaining robustness even when predictions are inaccurate. For the ski rental problem, the authors present two algorithms: a deterministic one that is robust and consistent, and a randomized one that offers a better trade-off between robustness and consistency. The deterministic algorithm is (1+1/λ)-robust and (1+λ)-consistent, while the randomized algorithm is (1+1/b)/(1-e^{-(λ-1/b)})-robust and λ/(1-e^{-λ})-consistent. These algorithms are designed to be oblivious to the performance of the predictor, ensuring that they perform well even when predictions are poor. For non-clairvoyant job scheduling, the authors propose a randomized algorithm that is (2/(1-λ))-robust and (1/λ)-consistent. This algorithm combines the Shortest Predicted Job First (SPJF) algorithm with the round-robin algorithm to achieve better performance. The SPJF algorithm has a competitive ratio of (1 + 2η/n), where η is the total prediction error and n is the number of jobs. The preferential round-robin algorithm, which combines SPJF and round-robin, achieves a competitive ratio of min{1/λ(1 + 2η/n), 2/(1-λ)}. The authors also conduct experiments to evaluate the performance of their algorithms on real-world data. The results show that their algorithms outperform traditional algorithms, even when predictions are inaccurate. The experiments demonstrate that the algorithms are practical and achieve good performance compared to ones that do not use any prediction. The study highlights the potential of using ML predictions to improve the worst-case performance of online algorithms.This paper explores how machine learning (ML) predictions can be used to improve the performance of online algorithms. The authors study two classical online problems: ski rental and non-clairvoyant job scheduling. In the ski rental problem, a skier must decide whether to rent or buy skis based on the number of days they will ski. In non-clairvoyant job scheduling, the goal is to schedule jobs on a single machine with unknown processing times. The authors propose new online algorithms that use ML predictions to make decisions, achieving better performance than traditional algorithms while maintaining robustness even when predictions are inaccurate. For the ski rental problem, the authors present two algorithms: a deterministic one that is robust and consistent, and a randomized one that offers a better trade-off between robustness and consistency. The deterministic algorithm is (1+1/λ)-robust and (1+λ)-consistent, while the randomized algorithm is (1+1/b)/(1-e^{-(λ-1/b)})-robust and λ/(1-e^{-λ})-consistent. These algorithms are designed to be oblivious to the performance of the predictor, ensuring that they perform well even when predictions are poor. For non-clairvoyant job scheduling, the authors propose a randomized algorithm that is (2/(1-λ))-robust and (1/λ)-consistent. This algorithm combines the Shortest Predicted Job First (SPJF) algorithm with the round-robin algorithm to achieve better performance. The SPJF algorithm has a competitive ratio of (1 + 2η/n), where η is the total prediction error and n is the number of jobs. The preferential round-robin algorithm, which combines SPJF and round-robin, achieves a competitive ratio of min{1/λ(1 + 2η/n), 2/(1-λ)}. The authors also conduct experiments to evaluate the performance of their algorithms on real-world data. The results show that their algorithms outperform traditional algorithms, even when predictions are inaccurate. The experiments demonstrate that the algorithms are practical and achieve good performance compared to ones that do not use any prediction. The study highlights the potential of using ML predictions to improve the worst-case performance of online algorithms.