机器学习：送货司机的福音

尽管科技有了翻天覆地的发展，但许多公司为送货卡车规划路线的方式仍然和十年前一样。负责人前一天定好路线，并把打印的路线图交给司机，或是把它们传到司机裤子后面口袋放着的手机里。

不过如果司机在路上陷入堵车，他们只能怪运气不好，导致晚点。如果出现了意外的暴风雪导致道路无法通行，结果也是一样。

简而言之，这些路线太不灵活了。

不过波士顿的初创公司Wise Systems正在让机器学习来自司机手机中的数据，借此解决这个年代久远的问题。它会综合考虑驾驶速度、GPS定位以及包括交通路况、天气、订单目的地、客户收货时间等细节。

最后生成的就是可以根据任何状况随时调整的送货路线。如果这项技术判定司机会因为道路关闭等因素无法按计划抵达，就会调整全天的日程。如果无法调整，司机就会在手机上收到明显的警告，提醒需要加快送货速度。（红色说明这不是个好的提醒。）

这项技术旨在绘制路线，提高司机的工作效率。如此一来，公司可以通过提高司机一趟的送货量来节约成本，同时还能通过提高订单准点率或由客户偏好的司机运送的概率，从而取悦客户。

Wise Systems拥有15名员工，他们脱胎于麻省理工学院（MIT）创业学研究生班的一项任务。起初，公司四位创始人的计划是做犯罪地图，但是潜在的客户表示：“犯罪地图是个好主意，但是交通状况是更糟糕的事情。”来自麻省理工学院交通和物流中心（MIT Center for Transportation and Logistics）的团队指导教授也鼓励他们往这个方向发展。2014年，他们意识到这个想法蕴含潜力，于是成立了公司。

从本质上说，优化配送路径的问题就是数学家从1930年开始就试图解决的巡回售货员问题（Traveling Salesman Problem）。尽管任务简单易懂——让推销员以最高的效率来往于各城市，在最后返回家中——但目前为止答案仍然悬而未决。推销的路线有无穷多种可能性，就如送货一样。

Wise的机器学习就在此大显身手。它的算法会从日常数据中学习，从而改善技术提供的行驶路线。Wise的首席技术官阿里·卡米尔表示：“这项技术远远不只是获取数据并吸收，它还会忘记那些可能出错的路线。”

随着各公司都在努力争夺那些希望几乎立刻就收到货物的消费者，改善送货路线比以往更加重要。例如，亚马逊（Amazon）就给食品杂货和特定的Prime产品提供当日快递，一至两个小时内就能送到，这需要强大的运算能力和机器学习工具。

不过开发这样的系统难度很大。UPS十多年来一直在研发自己的客户软件Orion。据说有超过500人参与开发该项技术，不过十年的努力仍然无法让它得到完全应用。在曼哈顿，UPS的司机仍然在使用老版的ED系统，因为在复杂的城市环境下，Orion的表现并不理想。

2016年底，安海斯布希（Anheuser-Busch）同意在西雅图和圣迭戈的批发商那里试用软件，就此成为了Wise System的第一个客户。六个月后，这家啤酒巨头在美国更多的批发商处全面采用了这项技术——其中包括一个供司机使用的移动应用和一个供管理者使用的网络工具。今年2月，Wise把这项技术应用到了所有美国的批发商处——总共20家，另外还有两家位于加拿大的安大略和魁北克。

20多年来，安海斯布希使用的都是Roadnet，该技术可以生成当天的配送路线。Roadnet可以帮助建立计划，设置配送顺序，但是在司机上路之后，那些路线不再会发生变化。

当安海斯布希把Roadnet创立的路径与司机的实际行驶路径进行对比之后，一个明显的问题出现了。公司发现，司机往往会偏离计划。这表明司机认为自己比技术工具更了解实际。当他们多年来每天按照同样的路线行驶时，很容易出现这样的情况。它也凸显出一个问题：司机会积累一些路线相关的工作经验，但使用技术很难获取这种经验。

为了让系统整合这类经验，Wise Systems会让司机通过移动应用输入实时数据，例如客户是否希望由特定司机送货，某地是否难以停车等。这些共享的说明会通过代码加入应用，如此一来，算法在未来就能利用这一信息。当一个新司机开始使用现有路线时，这类共享的经验尤其能够起到帮助。

使用一年后，安海斯布希表示，他们发现Wise Systems的几大优势。公司批发运营主管马特洛克·罗杰斯表示：“Wise可以学习模式和历史，这让它在将来会效率更高。”Wise可以让他的团队看到司机的实时位置，这减少了电话沟通，也不必更新文字状态了。

安海斯布希表示，在员工得到培训，可以恰当使用该工具的城市市场，每站里程数减少了4%，这意味着燃料的节省，磨损率的降低，而工作效率的提高也让司机的收入水涨船高。

另一个好处在于客户服务的提升。过去，司机错过送货时段也不会收到提醒。罗杰斯表示：“如今，Wise会展示可能迟到的15个站点，如果我们要为特定的客户保证某些送货时段，还能把它们按优先级排列。”

Wise Systems表示，随着这项技术的用户越来越多，可供利用的数据越来越多，它还会渐渐改善。想象一下，如果Wise现有用户群的2,000名司机变成20,000名，他们的集体智慧将达到什么水平。

Wise的卡米尔表示：“在物流业，司机的薪水是与送货单数而不是工作时间挂钩的。所以他们天生就喜欢我们，因为我们可以帮助他们每天送出更多货物。”（财富中文网）

更正：本文之前错误地表示安海斯布希在美国的20家零售商和加拿大的另外两家零售商合作测试Wise Systems。实际上，那些测试是与批发商合作进行的。

译者：严匡正 

Despite radical advances in technology, many companies still plan routes for their delivery trucks the same way they did a decade ago. Managers create itineraries the day before, and then hand printouts to drivers to follow or add them to the hand-held devices that their drivers carry at their hip.

But when drivers get stuck in traffic jams while on their rounds, they’re simply out of luck and behind schedule. The same thing happens if there’s a surprise snowstorm that makes roads impassable.

In short, the routes are inflexible.

But Wise Systems, a Boston startup, is tackling this age-old problem by pairing machine learning with data it collects from drivers’ mobile phones. It crunches information like the driver’s speed and GPS location with other details including traffic, weather, where the order is being delivered, and when customers are available to receive their orders.

What emerges is a delivery route that can be tweaked on the fly depending on any complications that come up. If the technology determines that a driver will miss a scheduled stop because of road closures, for example, it will adjust the schedule for the entire day. If that’s not possible, the driver will receive alerts on his or her mobile phone as a not-so subtle hint to pick up the pace. (Red is not a good sign.)

The goal is to create routes that allow drivers to work more efficiently. By doing so, companies can save money by increasing the number of deliveries that drivers can make during shifts while also making customers happier by improving the likelihood that orders will arrive on time, or by the driver they prefer.

Wise Systems, which has 15 employees, grew out of an assignment in a graduate class on entrepreneurship at MIT. At first, the idea of the company’s four founders was to map crime, but potential customers told them “crime was good, but traffic is worse.” One of the teams’ advisors, from the MIT Center for Transportation and Logistics, also nudged them in this direction. In 2014, when they realized the idea had potential, the team incorporated the company.

Optimizing delivery routes has its roots in what’s called the Traveling Salesman Problem, which mathematicians have been trying to solve since 1930. While the task is straightforward––finding most efficient route between cities for salesmen before returning home––it remains unsolved. The possibilities are limitless, much like the possibilities for deliveries.

This is where Wise’s machine learning comes into play. Its algorithms learn from each day’s data so that it can improve the routes the technology provides going forward. “It’s much more than taking the data and feeding it in,” says Wise chief technology officer Ali Kamil. “It’s also unlearning some things that might go wrong.”

Increasing the efficiency of deliveries is now more important than ever for companies as they battle for customers who expect their orders almost immediately. Amazon, for example, offers same-day deliveries of groceries and certain Prime products within one- and two-hour delivery windows, requiring huge computing power and machine learning tools.

But creating such a system is difficult. UPS has been building its own custom software––Orion––for over a decade. Over 500 people reportedly worked on the technology, but after 10 years, it’s still not fully deployed. In Manhattan, UPS drivers still use an old version called ED because Orion doesn’t do well in complex urban environments.

Anheuser-Busch became Wise Systems’ first client when it agreed in late 2016 to test the software with its Seattle and San Diego wholesalers. Six months later, the beer giant rolled out the technology––a mobile app for drivers and a web-based tool for managers––to more of its’ wholesalers across the country. As of this week, Wise has been implemented at all of its U.S. wholesalers—20 in total, plus two others in Ontario and Québec

For over 20 years, Anheuser-Busch used Roadnet, a technology that creates delivery routes up to the day of. Roadnet helps build the plan and set the sequence, but those routes don’t change after drivers get on the road.

Another problem became apparent when Anheuser-Busch compared the routes the Roadnet software created with those that drivers actually took. The company found that drivers often deviated from the plan. It was a sign that drivers thought they knew better than the technology, an easy slip-up when they follow the same route every day for years. It also highlighted the problem of incorporating some of the on-the-job knowledge that drivers had about their routes that technology has difficulty capturing.

To get some of that expertise into its system, Wise Systems lets drivers enter real-time data through its mobile app. Examples include whether a customer prefers to be serviced by a specific driver and whether parking is scarce. These shared notes are added to the app with a code so that the algorithm can take that information into account in the future. This kind of shared knowledge can be especially helpful when a new driver takes over an existing route.

After one year, Anheuser-Busch says it’s noticed several benefits of using Wise Systems. “Wise learns patterns and history, which helps it be more effective in the future,” says Matlock Rogers, director of wholesale operations for Anheuser-Busch. It lets his team see where drivers are in real time, reducing the phone calls and texting otherwise required for updates.

In urban markets where employees are trained and using the tools properly, Anheuser-Busch says it has reduced the miles traveled per stop by 4%, which translates into fuel savings, lower wear and tear on trucks, and, for the driver, improved earnings based on higher productivity.

Another benefit is improved customer service. In the past, drivers wouldn’t be alerted to missing a delivery window. “Now, Wise will show us the last 15 stops might be late and we can prioritize them if we need to hit a specific window for a certain client,” says Rogers.

Wise Systems says that its technology will improve over time as it takes on more clients, which in turn provide its system with more data to crunch. Imagine a network of 2,000 drivers––Wise’s pool now––versus one that taps the collective brainpower of 20,000 drivers.

“In the logistics industry drivers are paid by deliveries made, not time,” says Wise’s Kamil. “Inherently, they love us because we help them make more deliveries in a day.”

Correction: An earlier version of this article mistakenly said that Anheuser-Busch was testing Wise Systems technology with 20 retailers in the U.S. and two others in Canada. In fact, those tests are with wholesalers.