Uber如何使用Ray®优化打车业务

Computational efficiency is a significant challenge when scaling solutions to a marketplace as large and as complex as Uber. The running and tuning of the Uber rides business relies on substantial numbers of machine learning models and optimization algorithms. Luckily, many pieces of the system can be treated in parallel. Ray® is a general compute engine for Python® that’s designed for ML, AI, and other algorithmic workloads. We describe how Uber has adopted Ray to enable mission-critical systems. Using Uber’s mobility marketplace allocation tuning system as an example, we found performance improvements of up to 40 times that unlocked new capabilities. It also improved developer productivity by increasing iteration speed, reducing incident mitigation time, and lowering code complexity.

当将解决方案扩展到像Uber这样大而复杂的市场时,计算效率是一个重大挑战。Uber打车业务的运行和调优依赖于大量的机器学习模型和优化算法。幸运的是,系统的许多部分可以并行处理。Ray®是一个为Python®设计的通用计算引擎,适用于ML、AI和其他算法工作负载。我们描述了Uber如何采用Ray来支持关键任务系统。以Uber的移动市场分配调优系统为例,我们发现性能提升了多达40倍,解锁了新的能力。它还通过提高迭代速度、减少事件缓解时间和降低代码复杂性来提高开发人员的生产力。

To manage the health and efficiency of the mobility marketplace, Uber has several levers it can adjust, such as offering incentives for drivers to complete a certain number of trips per week or promotions that provide riders with discounts on their trips. Figuring out how to set these levers to efficiently achieve various targets and maximize outcomes is both a technically challenging task and a high-value opportunity for the company.

为了管理移动市场的健康和效率,Uber有几个可以调整的杠杆,例如为司机提供每周完成一定数量行程的激励措施,或为乘客提供行程折扣的促销活动。如何设置这些杠杆以有效地实现各种目标并最大化结果,既是一个技术上具有挑战性的任务,也是公司的一次高价值机会。

One of the problems we’d like to solve is shown in Figure 1:

我们想要解决的问题之一如图1所示:

Image

Figure 1: Marketplace incentive objective function.

图1:市场激励目标函数。

In this problem, we’d like to maximize some objective “f” that has some value to the business by changing the variables we control, “b.” Since these variables are controlled per city, we refer to the granularity as city-lever. We’re also subject to some constraints that are meant to keep the marketplace ...

开通本站会员,查看完整译文。

Главная - Вики-сайт
Copyright © 2011-2025 iteam. Current version is 2.139.1. UTC+08:00, 2025-01-12 06:42
浙ICP备14020137号-1 $Гость$